2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
24 #include "ext4_jbd2.h"
26 #include <linux/log2.h>
27 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <trace/events/ext4.h>
31 #ifdef CONFIG_EXT4_DEBUG
32 ushort ext4_mballoc_debug __read_mostly;
34 module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
35 MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
40 * - test ext4_ext_search_left() and ext4_ext_search_right()
41 * - search for metadata in few groups
44 * - normalization should take into account whether file is still open
45 * - discard preallocations if no free space left (policy?)
46 * - don't normalize tails
48 * - reservation for superuser
51 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
52 * - track min/max extents in each group for better group selection
53 * - mb_mark_used() may allocate chunk right after splitting buddy
54 * - tree of groups sorted by number of free blocks
59 * The allocation request involve request for multiple number of blocks
60 * near to the goal(block) value specified.
62 * During initialization phase of the allocator we decide to use the
63 * group preallocation or inode preallocation depending on the size of
64 * the file. The size of the file could be the resulting file size we
65 * would have after allocation, or the current file size, which ever
66 * is larger. If the size is less than sbi->s_mb_stream_request we
67 * select to use the group preallocation. The default value of
68 * s_mb_stream_request is 16 blocks. This can also be tuned via
69 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
70 * terms of number of blocks.
72 * The main motivation for having small file use group preallocation is to
73 * ensure that we have small files closer together on the disk.
75 * First stage the allocator looks at the inode prealloc list,
76 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
77 * spaces for this particular inode. The inode prealloc space is
80 * pa_lstart -> the logical start block for this prealloc space
81 * pa_pstart -> the physical start block for this prealloc space
82 * pa_len -> length for this prealloc space (in clusters)
83 * pa_free -> free space available in this prealloc space (in clusters)
85 * The inode preallocation space is used looking at the _logical_ start
86 * block. If only the logical file block falls within the range of prealloc
87 * space we will consume the particular prealloc space. This makes sure that
88 * we have contiguous physical blocks representing the file blocks
90 * The important thing to be noted in case of inode prealloc space is that
91 * we don't modify the values associated to inode prealloc space except
94 * If we are not able to find blocks in the inode prealloc space and if we
95 * have the group allocation flag set then we look at the locality group
96 * prealloc space. These are per CPU prealloc list represented as
98 * ext4_sb_info.s_locality_groups[smp_processor_id()]
100 * The reason for having a per cpu locality group is to reduce the contention
101 * between CPUs. It is possible to get scheduled at this point.
103 * The locality group prealloc space is used looking at whether we have
104 * enough free space (pa_free) within the prealloc space.
106 * If we can't allocate blocks via inode prealloc or/and locality group
107 * prealloc then we look at the buddy cache. The buddy cache is represented
108 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
109 * mapped to the buddy and bitmap information regarding different
110 * groups. The buddy information is attached to buddy cache inode so that
111 * we can access them through the page cache. The information regarding
112 * each group is loaded via ext4_mb_load_buddy. The information involve
113 * block bitmap and buddy information. The information are stored in the
117 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
120 * one block each for bitmap and buddy information. So for each group we
121 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
122 * blocksize) blocks. So it can have information regarding groups_per_page
123 * which is blocks_per_page/2
125 * The buddy cache inode is not stored on disk. The inode is thrown
126 * away when the filesystem is unmounted.
128 * We look for count number of blocks in the buddy cache. If we were able
129 * to locate that many free blocks we return with additional information
130 * regarding rest of the contiguous physical block available
132 * Before allocating blocks via buddy cache we normalize the request
133 * blocks. This ensure we ask for more blocks that we needed. The extra
134 * blocks that we get after allocation is added to the respective prealloc
135 * list. In case of inode preallocation we follow a list of heuristics
136 * based on file size. This can be found in ext4_mb_normalize_request. If
137 * we are doing a group prealloc we try to normalize the request to
138 * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is
139 * dependent on the cluster size; for non-bigalloc file systems, it is
140 * 512 blocks. This can be tuned via
141 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
142 * terms of number of blocks. If we have mounted the file system with -O
143 * stripe=<value> option the group prealloc request is normalized to the
144 * the smallest multiple of the stripe value (sbi->s_stripe) which is
145 * greater than the default mb_group_prealloc.
147 * The regular allocator (using the buddy cache) supports a few tunables.
149 * /sys/fs/ext4/<partition>/mb_min_to_scan
150 * /sys/fs/ext4/<partition>/mb_max_to_scan
151 * /sys/fs/ext4/<partition>/mb_order2_req
153 * The regular allocator uses buddy scan only if the request len is power of
154 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
155 * value of s_mb_order2_reqs can be tuned via
156 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
157 * stripe size (sbi->s_stripe), we try to search for contiguous block in
158 * stripe size. This should result in better allocation on RAID setups. If
159 * not, we search in the specific group using bitmap for best extents. The
160 * tunable min_to_scan and max_to_scan control the behaviour here.
161 * min_to_scan indicate how long the mballoc __must__ look for a best
162 * extent and max_to_scan indicates how long the mballoc __can__ look for a
163 * best extent in the found extents. Searching for the blocks starts with
164 * the group specified as the goal value in allocation context via
165 * ac_g_ex. Each group is first checked based on the criteria whether it
166 * can be used for allocation. ext4_mb_good_group explains how the groups are
169 * Both the prealloc space are getting populated as above. So for the first
170 * request we will hit the buddy cache which will result in this prealloc
171 * space getting filled. The prealloc space is then later used for the
172 * subsequent request.
176 * mballoc operates on the following data:
178 * - in-core buddy (actually includes buddy and bitmap)
179 * - preallocation descriptors (PAs)
181 * there are two types of preallocations:
183 * assiged to specific inode and can be used for this inode only.
184 * it describes part of inode's space preallocated to specific
185 * physical blocks. any block from that preallocated can be used
186 * independent. the descriptor just tracks number of blocks left
187 * unused. so, before taking some block from descriptor, one must
188 * make sure corresponded logical block isn't allocated yet. this
189 * also means that freeing any block within descriptor's range
190 * must discard all preallocated blocks.
192 * assigned to specific locality group which does not translate to
193 * permanent set of inodes: inode can join and leave group. space
194 * from this type of preallocation can be used for any inode. thus
195 * it's consumed from the beginning to the end.
197 * relation between them can be expressed as:
198 * in-core buddy = on-disk bitmap + preallocation descriptors
200 * this mean blocks mballoc considers used are:
201 * - allocated blocks (persistent)
202 * - preallocated blocks (non-persistent)
204 * consistency in mballoc world means that at any time a block is either
205 * free or used in ALL structures. notice: "any time" should not be read
206 * literally -- time is discrete and delimited by locks.
208 * to keep it simple, we don't use block numbers, instead we count number of
209 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
211 * all operations can be expressed as:
212 * - init buddy: buddy = on-disk + PAs
213 * - new PA: buddy += N; PA = N
214 * - use inode PA: on-disk += N; PA -= N
215 * - discard inode PA buddy -= on-disk - PA; PA = 0
216 * - use locality group PA on-disk += N; PA -= N
217 * - discard locality group PA buddy -= PA; PA = 0
218 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
219 * is used in real operation because we can't know actual used
220 * bits from PA, only from on-disk bitmap
222 * if we follow this strict logic, then all operations above should be atomic.
223 * given some of them can block, we'd have to use something like semaphores
224 * killing performance on high-end SMP hardware. let's try to relax it using
225 * the following knowledge:
226 * 1) if buddy is referenced, it's already initialized
227 * 2) while block is used in buddy and the buddy is referenced,
228 * nobody can re-allocate that block
229 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
230 * bit set and PA claims same block, it's OK. IOW, one can set bit in
231 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
234 * so, now we're building a concurrency table:
237 * blocks for PA are allocated in the buddy, buddy must be referenced
238 * until PA is linked to allocation group to avoid concurrent buddy init
240 * we need to make sure that either on-disk bitmap or PA has uptodate data
241 * given (3) we care that PA-=N operation doesn't interfere with init
243 * the simplest way would be to have buddy initialized by the discard
244 * - use locality group PA
245 * again PA-=N must be serialized with init
246 * - discard locality group PA
247 * the simplest way would be to have buddy initialized by the discard
250 * i_data_sem serializes them
252 * discard process must wait until PA isn't used by another process
253 * - use locality group PA
254 * some mutex should serialize them
255 * - discard locality group PA
256 * discard process must wait until PA isn't used by another process
259 * i_data_sem or another mutex should serializes them
261 * discard process must wait until PA isn't used by another process
262 * - use locality group PA
263 * nothing wrong here -- they're different PAs covering different blocks
264 * - discard locality group PA
265 * discard process must wait until PA isn't used by another process
267 * now we're ready to make few consequences:
268 * - PA is referenced and while it is no discard is possible
269 * - PA is referenced until block isn't marked in on-disk bitmap
270 * - PA changes only after on-disk bitmap
271 * - discard must not compete with init. either init is done before
272 * any discard or they're serialized somehow
273 * - buddy init as sum of on-disk bitmap and PAs is done atomically
275 * a special case when we've used PA to emptiness. no need to modify buddy
276 * in this case, but we should care about concurrent init
281 * Logic in few words:
286 * mark bits in on-disk bitmap
289 * - use preallocation:
290 * find proper PA (per-inode or group)
292 * mark bits in on-disk bitmap
298 * mark bits in on-disk bitmap
301 * - discard preallocations in group:
303 * move them onto local list
304 * load on-disk bitmap
306 * remove PA from object (inode or locality group)
307 * mark free blocks in-core
309 * - discard inode's preallocations:
316 * - bitlock on a group (group)
317 * - object (inode/locality) (object)
328 * - release consumed pa:
333 * - generate in-core bitmap:
337 * - discard all for given object (inode, locality group):
342 * - discard all for given group:
349 static struct kmem_cache *ext4_pspace_cachep;
350 static struct kmem_cache *ext4_ac_cachep;
351 static struct kmem_cache *ext4_free_data_cachep;
353 /* We create slab caches for groupinfo data structures based on the
354 * superblock block size. There will be one per mounted filesystem for
355 * each unique s_blocksize_bits */
356 #define NR_GRPINFO_CACHES 8
357 static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
359 static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
360 "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
361 "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
362 "ext4_groupinfo_64k", "ext4_groupinfo_128k"
365 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
367 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
369 static void ext4_free_data_callback(struct super_block *sb,
370 struct ext4_journal_cb_entry *jce, int rc);
372 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
374 #if BITS_PER_LONG == 64
375 *bit += ((unsigned long) addr & 7UL) << 3;
376 addr = (void *) ((unsigned long) addr & ~7UL);
377 #elif BITS_PER_LONG == 32
378 *bit += ((unsigned long) addr & 3UL) << 3;
379 addr = (void *) ((unsigned long) addr & ~3UL);
381 #error "how many bits you are?!"
386 static inline int mb_test_bit(int bit, void *addr)
389 * ext4_test_bit on architecture like powerpc
390 * needs unsigned long aligned address
392 addr = mb_correct_addr_and_bit(&bit, addr);
393 return ext4_test_bit(bit, addr);
396 static inline void mb_set_bit(int bit, void *addr)
398 addr = mb_correct_addr_and_bit(&bit, addr);
399 ext4_set_bit(bit, addr);
402 static inline void mb_clear_bit(int bit, void *addr)
404 addr = mb_correct_addr_and_bit(&bit, addr);
405 ext4_clear_bit(bit, addr);
408 static inline int mb_test_and_clear_bit(int bit, void *addr)
410 addr = mb_correct_addr_and_bit(&bit, addr);
411 return ext4_test_and_clear_bit(bit, addr);
414 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
416 int fix = 0, ret, tmpmax;
417 addr = mb_correct_addr_and_bit(&fix, addr);
421 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
427 static inline int mb_find_next_bit(void *addr, int max, int start)
429 int fix = 0, ret, tmpmax;
430 addr = mb_correct_addr_and_bit(&fix, addr);
434 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
440 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
444 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
447 if (order > e4b->bd_blkbits + 1) {
452 /* at order 0 we see each particular block */
454 *max = 1 << (e4b->bd_blkbits + 3);
455 return e4b->bd_bitmap;
458 bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
459 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
465 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
466 int first, int count)
469 struct super_block *sb = e4b->bd_sb;
471 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
473 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
474 for (i = 0; i < count; i++) {
475 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
476 ext4_fsblk_t blocknr;
478 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
479 blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
480 ext4_grp_locked_error(sb, e4b->bd_group,
481 inode ? inode->i_ino : 0,
483 "freeing block already freed "
487 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
491 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
495 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
497 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
498 for (i = 0; i < count; i++) {
499 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
500 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
504 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
506 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
507 unsigned char *b1, *b2;
509 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
510 b2 = (unsigned char *) bitmap;
511 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
512 if (b1[i] != b2[i]) {
513 ext4_msg(e4b->bd_sb, KERN_ERR,
514 "corruption in group %u "
515 "at byte %u(%u): %x in copy != %x "
517 e4b->bd_group, i, i * 8, b1[i], b2[i]);
525 static inline void mb_free_blocks_double(struct inode *inode,
526 struct ext4_buddy *e4b, int first, int count)
530 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
531 int first, int count)
535 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
541 #ifdef AGGRESSIVE_CHECK
543 #define MB_CHECK_ASSERT(assert) \
547 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
548 function, file, line, # assert); \
553 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
554 const char *function, int line)
556 struct super_block *sb = e4b->bd_sb;
557 int order = e4b->bd_blkbits + 1;
564 struct ext4_group_info *grp;
567 struct list_head *cur;
572 static int mb_check_counter;
573 if (mb_check_counter++ % 100 != 0)
578 buddy = mb_find_buddy(e4b, order, &max);
579 MB_CHECK_ASSERT(buddy);
580 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
581 MB_CHECK_ASSERT(buddy2);
582 MB_CHECK_ASSERT(buddy != buddy2);
583 MB_CHECK_ASSERT(max * 2 == max2);
586 for (i = 0; i < max; i++) {
588 if (mb_test_bit(i, buddy)) {
589 /* only single bit in buddy2 may be 1 */
590 if (!mb_test_bit(i << 1, buddy2)) {
592 mb_test_bit((i<<1)+1, buddy2));
593 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
595 mb_test_bit(i << 1, buddy2));
600 /* both bits in buddy2 must be 1 */
601 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
602 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
604 for (j = 0; j < (1 << order); j++) {
605 k = (i * (1 << order)) + j;
607 !mb_test_bit(k, e4b->bd_bitmap));
611 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
616 buddy = mb_find_buddy(e4b, 0, &max);
617 for (i = 0; i < max; i++) {
618 if (!mb_test_bit(i, buddy)) {
619 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
627 /* check used bits only */
628 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
629 buddy2 = mb_find_buddy(e4b, j, &max2);
631 MB_CHECK_ASSERT(k < max2);
632 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
635 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
636 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
638 grp = ext4_get_group_info(sb, e4b->bd_group);
639 list_for_each(cur, &grp->bb_prealloc_list) {
640 ext4_group_t groupnr;
641 struct ext4_prealloc_space *pa;
642 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
643 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
644 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
645 for (i = 0; i < pa->pa_len; i++)
646 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
650 #undef MB_CHECK_ASSERT
651 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
652 __FILE__, __func__, __LINE__)
654 #define mb_check_buddy(e4b)
658 * Divide blocks started from @first with length @len into
659 * smaller chunks with power of 2 blocks.
660 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
661 * then increase bb_counters[] for corresponded chunk size.
663 static void ext4_mb_mark_free_simple(struct super_block *sb,
664 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
665 struct ext4_group_info *grp)
667 struct ext4_sb_info *sbi = EXT4_SB(sb);
671 unsigned short border;
673 BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
675 border = 2 << sb->s_blocksize_bits;
678 /* find how many blocks can be covered since this position */
679 max = ffs(first | border) - 1;
681 /* find how many blocks of power 2 we need to mark */
688 /* mark multiblock chunks only */
689 grp->bb_counters[min]++;
691 mb_clear_bit(first >> min,
692 buddy + sbi->s_mb_offsets[min]);
700 * Cache the order of the largest free extent we have available in this block
704 mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
709 grp->bb_largest_free_order = -1; /* uninit */
711 bits = sb->s_blocksize_bits + 1;
712 for (i = bits; i >= 0; i--) {
713 if (grp->bb_counters[i] > 0) {
714 grp->bb_largest_free_order = i;
720 static noinline_for_stack
721 void ext4_mb_generate_buddy(struct super_block *sb,
722 void *buddy, void *bitmap, ext4_group_t group)
724 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
725 ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
730 unsigned fragments = 0;
731 unsigned long long period = get_cycles();
733 /* initialize buddy from bitmap which is aggregation
734 * of on-disk bitmap and preallocations */
735 i = mb_find_next_zero_bit(bitmap, max, 0);
736 grp->bb_first_free = i;
740 i = mb_find_next_bit(bitmap, max, i);
744 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
746 grp->bb_counters[0]++;
748 i = mb_find_next_zero_bit(bitmap, max, i);
750 grp->bb_fragments = fragments;
752 if (free != grp->bb_free) {
753 ext4_grp_locked_error(sb, group, 0, 0,
754 "block bitmap and bg descriptor "
755 "inconsistent: %u vs %u free clusters",
758 * If we intend to continue, we consider group descriptor
759 * corrupt and update bb_free using bitmap value
762 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
764 mb_set_largest_free_order(sb, grp);
766 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
768 period = get_cycles() - period;
769 spin_lock(&EXT4_SB(sb)->s_bal_lock);
770 EXT4_SB(sb)->s_mb_buddies_generated++;
771 EXT4_SB(sb)->s_mb_generation_time += period;
772 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
775 static void mb_regenerate_buddy(struct ext4_buddy *e4b)
781 while ((buddy = mb_find_buddy(e4b, order++, &count))) {
782 ext4_set_bits(buddy, 0, count);
784 e4b->bd_info->bb_fragments = 0;
785 memset(e4b->bd_info->bb_counters, 0,
786 sizeof(*e4b->bd_info->bb_counters) *
787 (e4b->bd_sb->s_blocksize_bits + 2));
789 ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
790 e4b->bd_bitmap, e4b->bd_group);
793 /* The buddy information is attached the buddy cache inode
794 * for convenience. The information regarding each group
795 * is loaded via ext4_mb_load_buddy. The information involve
796 * block bitmap and buddy information. The information are
797 * stored in the inode as
800 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
803 * one block each for bitmap and buddy information.
804 * So for each group we take up 2 blocks. A page can
805 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
806 * So it can have information regarding groups_per_page which
807 * is blocks_per_page/2
809 * Locking note: This routine takes the block group lock of all groups
810 * for this page; do not hold this lock when calling this routine!
813 static int ext4_mb_init_cache(struct page *page, char *incore)
815 ext4_group_t ngroups;
821 ext4_group_t first_group, group;
823 struct super_block *sb;
824 struct buffer_head *bhs;
825 struct buffer_head **bh = NULL;
829 struct ext4_group_info *grinfo;
831 mb_debug(1, "init page %lu\n", page->index);
833 inode = page->mapping->host;
835 ngroups = ext4_get_groups_count(sb);
836 blocksize = 1 << inode->i_blkbits;
837 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
839 groups_per_page = blocks_per_page >> 1;
840 if (groups_per_page == 0)
843 /* allocate buffer_heads to read bitmaps */
844 if (groups_per_page > 1) {
845 i = sizeof(struct buffer_head *) * groups_per_page;
846 bh = kzalloc(i, GFP_NOFS);
854 first_group = page->index * blocks_per_page / 2;
856 /* read all groups the page covers into the cache */
857 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
858 if (group >= ngroups)
861 grinfo = ext4_get_group_info(sb, group);
863 * If page is uptodate then we came here after online resize
864 * which added some new uninitialized group info structs, so
865 * we must skip all initialized uptodate buddies on the page,
866 * which may be currently in use by an allocating task.
868 if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
872 if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
876 mb_debug(1, "read bitmap for group %u\n", group);
879 /* wait for I/O completion */
880 for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
881 if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
887 first_block = page->index * blocks_per_page;
888 for (i = 0; i < blocks_per_page; i++) {
889 group = (first_block + i) >> 1;
890 if (group >= ngroups)
893 if (!bh[group - first_group])
894 /* skip initialized uptodate buddy */
898 * data carry information regarding this
899 * particular group in the format specified
903 data = page_address(page) + (i * blocksize);
904 bitmap = bh[group - first_group]->b_data;
907 * We place the buddy block and bitmap block
910 if ((first_block + i) & 1) {
911 /* this is block of buddy */
912 BUG_ON(incore == NULL);
913 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
914 group, page->index, i * blocksize);
915 trace_ext4_mb_buddy_bitmap_load(sb, group);
916 grinfo = ext4_get_group_info(sb, group);
917 grinfo->bb_fragments = 0;
918 memset(grinfo->bb_counters, 0,
919 sizeof(*grinfo->bb_counters) *
920 (sb->s_blocksize_bits+2));
922 * incore got set to the group block bitmap below
924 ext4_lock_group(sb, group);
926 memset(data, 0xff, blocksize);
927 ext4_mb_generate_buddy(sb, data, incore, group);
928 ext4_unlock_group(sb, group);
931 /* this is block of bitmap */
932 BUG_ON(incore != NULL);
933 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
934 group, page->index, i * blocksize);
935 trace_ext4_mb_bitmap_load(sb, group);
937 /* see comments in ext4_mb_put_pa() */
938 ext4_lock_group(sb, group);
939 memcpy(data, bitmap, blocksize);
941 /* mark all preallocated blks used in in-core bitmap */
942 ext4_mb_generate_from_pa(sb, data, group);
943 ext4_mb_generate_from_freelist(sb, data, group);
944 ext4_unlock_group(sb, group);
946 /* set incore so that the buddy information can be
947 * generated using this
952 SetPageUptodate(page);
956 for (i = 0; i < groups_per_page; i++)
965 * Lock the buddy and bitmap pages. This make sure other parallel init_group
966 * on the same buddy page doesn't happen whild holding the buddy page lock.
967 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
968 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
970 static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
971 ext4_group_t group, struct ext4_buddy *e4b)
973 struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
974 int block, pnum, poff;
978 e4b->bd_buddy_page = NULL;
979 e4b->bd_bitmap_page = NULL;
981 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
983 * the buddy cache inode stores the block bitmap
984 * and buddy information in consecutive blocks.
985 * So for each group we need two blocks.
988 pnum = block / blocks_per_page;
989 poff = block % blocks_per_page;
990 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
993 BUG_ON(page->mapping != inode->i_mapping);
994 e4b->bd_bitmap_page = page;
995 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
997 if (blocks_per_page >= 2) {
998 /* buddy and bitmap are on the same page */
1003 pnum = block / blocks_per_page;
1004 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1007 BUG_ON(page->mapping != inode->i_mapping);
1008 e4b->bd_buddy_page = page;
1012 static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1014 if (e4b->bd_bitmap_page) {
1015 unlock_page(e4b->bd_bitmap_page);
1016 page_cache_release(e4b->bd_bitmap_page);
1018 if (e4b->bd_buddy_page) {
1019 unlock_page(e4b->bd_buddy_page);
1020 page_cache_release(e4b->bd_buddy_page);
1025 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1026 * block group lock of all groups for this page; do not hold the BG lock when
1027 * calling this routine!
1029 static noinline_for_stack
1030 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1033 struct ext4_group_info *this_grp;
1034 struct ext4_buddy e4b;
1039 mb_debug(1, "init group %u\n", group);
1040 this_grp = ext4_get_group_info(sb, group);
1042 * This ensures that we don't reinit the buddy cache
1043 * page which map to the group from which we are already
1044 * allocating. If we are looking at the buddy cache we would
1045 * have taken a reference using ext4_mb_load_buddy and that
1046 * would have pinned buddy page to page cache.
1048 ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1049 if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1051 * somebody initialized the group
1052 * return without doing anything
1057 page = e4b.bd_bitmap_page;
1058 ret = ext4_mb_init_cache(page, NULL);
1061 if (!PageUptodate(page)) {
1065 mark_page_accessed(page);
1067 if (e4b.bd_buddy_page == NULL) {
1069 * If both the bitmap and buddy are in
1070 * the same page we don't need to force
1076 /* init buddy cache */
1077 page = e4b.bd_buddy_page;
1078 ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1081 if (!PageUptodate(page)) {
1085 mark_page_accessed(page);
1087 ext4_mb_put_buddy_page_lock(&e4b);
1092 * Locking note: This routine calls ext4_mb_init_cache(), which takes the
1093 * block group lock of all groups for this page; do not hold the BG lock when
1094 * calling this routine!
1096 static noinline_for_stack int
1097 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1098 struct ext4_buddy *e4b)
1100 int blocks_per_page;
1106 struct ext4_group_info *grp;
1107 struct ext4_sb_info *sbi = EXT4_SB(sb);
1108 struct inode *inode = sbi->s_buddy_cache;
1111 mb_debug(1, "load group %u\n", group);
1113 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1114 grp = ext4_get_group_info(sb, group);
1116 e4b->bd_blkbits = sb->s_blocksize_bits;
1119 e4b->bd_group = group;
1120 e4b->bd_buddy_page = NULL;
1121 e4b->bd_bitmap_page = NULL;
1123 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1125 * we need full data about the group
1126 * to make a good selection
1128 ret = ext4_mb_init_group(sb, group);
1134 * the buddy cache inode stores the block bitmap
1135 * and buddy information in consecutive blocks.
1136 * So for each group we need two blocks.
1139 pnum = block / blocks_per_page;
1140 poff = block % blocks_per_page;
1142 /* we could use find_or_create_page(), but it locks page
1143 * what we'd like to avoid in fast path ... */
1144 page = find_get_page(inode->i_mapping, pnum);
1145 if (page == NULL || !PageUptodate(page)) {
1148 * drop the page reference and try
1149 * to get the page with lock. If we
1150 * are not uptodate that implies
1151 * somebody just created the page but
1152 * is yet to initialize the same. So
1153 * wait for it to initialize.
1155 page_cache_release(page);
1156 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1158 BUG_ON(page->mapping != inode->i_mapping);
1159 if (!PageUptodate(page)) {
1160 ret = ext4_mb_init_cache(page, NULL);
1165 mb_cmp_bitmaps(e4b, page_address(page) +
1166 (poff * sb->s_blocksize));
1171 if (page == NULL || !PageUptodate(page)) {
1175 e4b->bd_bitmap_page = page;
1176 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1177 mark_page_accessed(page);
1180 pnum = block / blocks_per_page;
1181 poff = block % blocks_per_page;
1183 page = find_get_page(inode->i_mapping, pnum);
1184 if (page == NULL || !PageUptodate(page)) {
1186 page_cache_release(page);
1187 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1189 BUG_ON(page->mapping != inode->i_mapping);
1190 if (!PageUptodate(page)) {
1191 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1200 if (page == NULL || !PageUptodate(page)) {
1204 e4b->bd_buddy_page = page;
1205 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1206 mark_page_accessed(page);
1208 BUG_ON(e4b->bd_bitmap_page == NULL);
1209 BUG_ON(e4b->bd_buddy_page == NULL);
1215 page_cache_release(page);
1216 if (e4b->bd_bitmap_page)
1217 page_cache_release(e4b->bd_bitmap_page);
1218 if (e4b->bd_buddy_page)
1219 page_cache_release(e4b->bd_buddy_page);
1220 e4b->bd_buddy = NULL;
1221 e4b->bd_bitmap = NULL;
1225 static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1227 if (e4b->bd_bitmap_page)
1228 page_cache_release(e4b->bd_bitmap_page);
1229 if (e4b->bd_buddy_page)
1230 page_cache_release(e4b->bd_buddy_page);
1234 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1239 BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1240 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1243 while (order <= e4b->bd_blkbits + 1) {
1245 if (!mb_test_bit(block, bb)) {
1246 /* this block is part of buddy of order 'order' */
1249 bb += 1 << (e4b->bd_blkbits - order);
1255 static void mb_clear_bits(void *bm, int cur, int len)
1261 if ((cur & 31) == 0 && (len - cur) >= 32) {
1262 /* fast path: clear whole word at once */
1263 addr = bm + (cur >> 3);
1268 mb_clear_bit(cur, bm);
1273 /* clear bits in given range
1274 * will return first found zero bit if any, -1 otherwise
1276 static int mb_test_and_clear_bits(void *bm, int cur, int len)
1283 if ((cur & 31) == 0 && (len - cur) >= 32) {
1284 /* fast path: clear whole word at once */
1285 addr = bm + (cur >> 3);
1286 if (*addr != (__u32)(-1) && zero_bit == -1)
1287 zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1292 if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1300 void ext4_set_bits(void *bm, int cur, int len)
1306 if ((cur & 31) == 0 && (len - cur) >= 32) {
1307 /* fast path: set whole word at once */
1308 addr = bm + (cur >> 3);
1313 mb_set_bit(cur, bm);
1319 * _________________________________________________________________ */
1321 static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1323 if (mb_test_bit(*bit + side, bitmap)) {
1324 mb_clear_bit(*bit, bitmap);
1330 mb_set_bit(*bit, bitmap);
1335 static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1339 void *buddy = mb_find_buddy(e4b, order, &max);
1344 /* Bits in range [first; last] are known to be set since
1345 * corresponding blocks were allocated. Bits in range
1346 * (first; last) will stay set because they form buddies on
1347 * upper layer. We just deal with borders if they don't
1348 * align with upper layer and then go up.
1349 * Releasing entire group is all about clearing
1350 * single bit of highest order buddy.
1354 * ---------------------------------
1356 * ---------------------------------
1357 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1358 * ---------------------------------
1360 * \_____________________/
1362 * Neither [1] nor [6] is aligned to above layer.
1363 * Left neighbour [0] is free, so mark it busy,
1364 * decrease bb_counters and extend range to
1366 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1367 * mark [6] free, increase bb_counters and shrink range to
1369 * Then shift range to [0; 2], go up and do the same.
1374 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1376 e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1381 if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1382 mb_clear_bits(buddy, first, last - first + 1);
1383 e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1392 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1393 int first, int count)
1395 int left_is_free = 0;
1396 int right_is_free = 0;
1398 int last = first + count - 1;
1399 struct super_block *sb = e4b->bd_sb;
1401 if (WARN_ON(count == 0))
1403 BUG_ON(last >= (sb->s_blocksize << 3));
1404 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1405 /* Don't bother if the block group is corrupt. */
1406 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1409 mb_check_buddy(e4b);
1410 mb_free_blocks_double(inode, e4b, first, count);
1412 e4b->bd_info->bb_free += count;
1413 if (first < e4b->bd_info->bb_first_free)
1414 e4b->bd_info->bb_first_free = first;
1416 /* access memory sequentially: check left neighbour,
1417 * clear range and then check right neighbour
1420 left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1421 block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1422 if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1423 right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1425 if (unlikely(block != -1)) {
1426 ext4_fsblk_t blocknr;
1428 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1429 blocknr += EXT4_C2B(EXT4_SB(sb), block);
1430 ext4_grp_locked_error(sb, e4b->bd_group,
1431 inode ? inode->i_ino : 0,
1433 "freeing already freed block "
1434 "(bit %u); block bitmap corrupt.",
1436 /* Mark the block group as corrupt. */
1437 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1438 &e4b->bd_info->bb_state);
1439 mb_regenerate_buddy(e4b);
1443 /* let's maintain fragments counter */
1444 if (left_is_free && right_is_free)
1445 e4b->bd_info->bb_fragments--;
1446 else if (!left_is_free && !right_is_free)
1447 e4b->bd_info->bb_fragments++;
1449 /* buddy[0] == bd_bitmap is a special case, so handle
1450 * it right away and let mb_buddy_mark_free stay free of
1451 * zero order checks.
1452 * Check if neighbours are to be coaleasced,
1453 * adjust bitmap bb_counters and borders appropriately.
1456 first += !left_is_free;
1457 e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1460 last -= !right_is_free;
1461 e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1465 mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1468 mb_set_largest_free_order(sb, e4b->bd_info);
1469 mb_check_buddy(e4b);
1472 static int mb_find_extent(struct ext4_buddy *e4b, int block,
1473 int needed, struct ext4_free_extent *ex)
1479 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1482 buddy = mb_find_buddy(e4b, 0, &max);
1483 BUG_ON(buddy == NULL);
1484 BUG_ON(block >= max);
1485 if (mb_test_bit(block, buddy)) {
1492 /* find actual order */
1493 order = mb_find_order_for_block(e4b, block);
1494 block = block >> order;
1496 ex->fe_len = 1 << order;
1497 ex->fe_start = block << order;
1498 ex->fe_group = e4b->bd_group;
1500 /* calc difference from given start */
1501 next = next - ex->fe_start;
1503 ex->fe_start += next;
1505 while (needed > ex->fe_len &&
1506 mb_find_buddy(e4b, order, &max)) {
1508 if (block + 1 >= max)
1511 next = (block + 1) * (1 << order);
1512 if (mb_test_bit(next, e4b->bd_bitmap))
1515 order = mb_find_order_for_block(e4b, next);
1517 block = next >> order;
1518 ex->fe_len += 1 << order;
1521 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1525 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1531 int start = ex->fe_start;
1532 int len = ex->fe_len;
1537 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1538 BUG_ON(e4b->bd_group != ex->fe_group);
1539 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1540 mb_check_buddy(e4b);
1541 mb_mark_used_double(e4b, start, len);
1543 e4b->bd_info->bb_free -= len;
1544 if (e4b->bd_info->bb_first_free == start)
1545 e4b->bd_info->bb_first_free += len;
1547 /* let's maintain fragments counter */
1549 mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1550 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1551 max = !mb_test_bit(start + len, e4b->bd_bitmap);
1553 e4b->bd_info->bb_fragments++;
1554 else if (!mlen && !max)
1555 e4b->bd_info->bb_fragments--;
1557 /* let's maintain buddy itself */
1559 ord = mb_find_order_for_block(e4b, start);
1561 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1562 /* the whole chunk may be allocated at once! */
1564 buddy = mb_find_buddy(e4b, ord, &max);
1565 BUG_ON((start >> ord) >= max);
1566 mb_set_bit(start >> ord, buddy);
1567 e4b->bd_info->bb_counters[ord]--;
1574 /* store for history */
1576 ret = len | (ord << 16);
1578 /* we have to split large buddy */
1580 buddy = mb_find_buddy(e4b, ord, &max);
1581 mb_set_bit(start >> ord, buddy);
1582 e4b->bd_info->bb_counters[ord]--;
1585 cur = (start >> ord) & ~1U;
1586 buddy = mb_find_buddy(e4b, ord, &max);
1587 mb_clear_bit(cur, buddy);
1588 mb_clear_bit(cur + 1, buddy);
1589 e4b->bd_info->bb_counters[ord]++;
1590 e4b->bd_info->bb_counters[ord]++;
1592 mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1594 ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1595 mb_check_buddy(e4b);
1601 * Must be called under group lock!
1603 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1604 struct ext4_buddy *e4b)
1606 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1609 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1610 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1612 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1613 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1614 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1616 /* preallocation can change ac_b_ex, thus we store actually
1617 * allocated blocks for history */
1618 ac->ac_f_ex = ac->ac_b_ex;
1620 ac->ac_status = AC_STATUS_FOUND;
1621 ac->ac_tail = ret & 0xffff;
1622 ac->ac_buddy = ret >> 16;
1625 * take the page reference. We want the page to be pinned
1626 * so that we don't get a ext4_mb_init_cache_call for this
1627 * group until we update the bitmap. That would mean we
1628 * double allocate blocks. The reference is dropped
1629 * in ext4_mb_release_context
1631 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1632 get_page(ac->ac_bitmap_page);
1633 ac->ac_buddy_page = e4b->bd_buddy_page;
1634 get_page(ac->ac_buddy_page);
1635 /* store last allocated for subsequent stream allocation */
1636 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1637 spin_lock(&sbi->s_md_lock);
1638 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1639 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1640 spin_unlock(&sbi->s_md_lock);
1645 * regular allocator, for general purposes allocation
1648 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1649 struct ext4_buddy *e4b,
1652 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1653 struct ext4_free_extent *bex = &ac->ac_b_ex;
1654 struct ext4_free_extent *gex = &ac->ac_g_ex;
1655 struct ext4_free_extent ex;
1658 if (ac->ac_status == AC_STATUS_FOUND)
1661 * We don't want to scan for a whole year
1663 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1664 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1665 ac->ac_status = AC_STATUS_BREAK;
1670 * Haven't found good chunk so far, let's continue
1672 if (bex->fe_len < gex->fe_len)
1675 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1676 && bex->fe_group == e4b->bd_group) {
1677 /* recheck chunk's availability - we don't know
1678 * when it was found (within this lock-unlock
1680 max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1681 if (max >= gex->fe_len) {
1682 ext4_mb_use_best_found(ac, e4b);
1689 * The routine checks whether found extent is good enough. If it is,
1690 * then the extent gets marked used and flag is set to the context
1691 * to stop scanning. Otherwise, the extent is compared with the
1692 * previous found extent and if new one is better, then it's stored
1693 * in the context. Later, the best found extent will be used, if
1694 * mballoc can't find good enough extent.
1696 * FIXME: real allocation policy is to be designed yet!
1698 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1699 struct ext4_free_extent *ex,
1700 struct ext4_buddy *e4b)
1702 struct ext4_free_extent *bex = &ac->ac_b_ex;
1703 struct ext4_free_extent *gex = &ac->ac_g_ex;
1705 BUG_ON(ex->fe_len <= 0);
1706 BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1707 BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1708 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1713 * The special case - take what you catch first
1715 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1717 ext4_mb_use_best_found(ac, e4b);
1722 * Let's check whether the chuck is good enough
1724 if (ex->fe_len == gex->fe_len) {
1726 ext4_mb_use_best_found(ac, e4b);
1731 * If this is first found extent, just store it in the context
1733 if (bex->fe_len == 0) {
1739 * If new found extent is better, store it in the context
1741 if (bex->fe_len < gex->fe_len) {
1742 /* if the request isn't satisfied, any found extent
1743 * larger than previous best one is better */
1744 if (ex->fe_len > bex->fe_len)
1746 } else if (ex->fe_len > gex->fe_len) {
1747 /* if the request is satisfied, then we try to find
1748 * an extent that still satisfy the request, but is
1749 * smaller than previous one */
1750 if (ex->fe_len < bex->fe_len)
1754 ext4_mb_check_limits(ac, e4b, 0);
1757 static noinline_for_stack
1758 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1759 struct ext4_buddy *e4b)
1761 struct ext4_free_extent ex = ac->ac_b_ex;
1762 ext4_group_t group = ex.fe_group;
1766 BUG_ON(ex.fe_len <= 0);
1767 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1771 ext4_lock_group(ac->ac_sb, group);
1772 max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1776 ext4_mb_use_best_found(ac, e4b);
1779 ext4_unlock_group(ac->ac_sb, group);
1780 ext4_mb_unload_buddy(e4b);
1785 static noinline_for_stack
1786 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1787 struct ext4_buddy *e4b)
1789 ext4_group_t group = ac->ac_g_ex.fe_group;
1792 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1793 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1794 struct ext4_free_extent ex;
1796 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1798 if (grp->bb_free == 0)
1801 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1805 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1806 ext4_mb_unload_buddy(e4b);
1810 ext4_lock_group(ac->ac_sb, group);
1811 max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1812 ac->ac_g_ex.fe_len, &ex);
1814 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1817 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1819 /* use do_div to get remainder (would be 64-bit modulo) */
1820 if (do_div(start, sbi->s_stripe) == 0) {
1823 ext4_mb_use_best_found(ac, e4b);
1825 } else if (max >= ac->ac_g_ex.fe_len) {
1826 BUG_ON(ex.fe_len <= 0);
1827 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1828 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1831 ext4_mb_use_best_found(ac, e4b);
1832 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1833 /* Sometimes, caller may want to merge even small
1834 * number of blocks to an existing extent */
1835 BUG_ON(ex.fe_len <= 0);
1836 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1837 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1840 ext4_mb_use_best_found(ac, e4b);
1842 ext4_unlock_group(ac->ac_sb, group);
1843 ext4_mb_unload_buddy(e4b);
1849 * The routine scans buddy structures (not bitmap!) from given order
1850 * to max order and tries to find big enough chunk to satisfy the req
1852 static noinline_for_stack
1853 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1854 struct ext4_buddy *e4b)
1856 struct super_block *sb = ac->ac_sb;
1857 struct ext4_group_info *grp = e4b->bd_info;
1863 BUG_ON(ac->ac_2order <= 0);
1864 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1865 if (grp->bb_counters[i] == 0)
1868 buddy = mb_find_buddy(e4b, i, &max);
1869 BUG_ON(buddy == NULL);
1871 k = mb_find_next_zero_bit(buddy, max, 0);
1876 ac->ac_b_ex.fe_len = 1 << i;
1877 ac->ac_b_ex.fe_start = k << i;
1878 ac->ac_b_ex.fe_group = e4b->bd_group;
1880 ext4_mb_use_best_found(ac, e4b);
1882 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1884 if (EXT4_SB(sb)->s_mb_stats)
1885 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1892 * The routine scans the group and measures all found extents.
1893 * In order to optimize scanning, caller must pass number of
1894 * free blocks in the group, so the routine can know upper limit.
1896 static noinline_for_stack
1897 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1898 struct ext4_buddy *e4b)
1900 struct super_block *sb = ac->ac_sb;
1901 void *bitmap = e4b->bd_bitmap;
1902 struct ext4_free_extent ex;
1906 free = e4b->bd_info->bb_free;
1909 i = e4b->bd_info->bb_first_free;
1911 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1912 i = mb_find_next_zero_bit(bitmap,
1913 EXT4_CLUSTERS_PER_GROUP(sb), i);
1914 if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1916 * IF we have corrupt bitmap, we won't find any
1917 * free blocks even though group info says we
1918 * we have free blocks
1920 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1921 "%d free clusters as per "
1922 "group info. But bitmap says 0",
1927 mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1928 BUG_ON(ex.fe_len <= 0);
1929 if (free < ex.fe_len) {
1930 ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1931 "%d free clusters as per "
1932 "group info. But got %d blocks",
1935 * The number of free blocks differs. This mostly
1936 * indicate that the bitmap is corrupt. So exit
1937 * without claiming the space.
1942 ext4_mb_measure_extent(ac, &ex, e4b);
1948 ext4_mb_check_limits(ac, e4b, 1);
1952 * This is a special case for storages like raid5
1953 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1955 static noinline_for_stack
1956 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1957 struct ext4_buddy *e4b)
1959 struct super_block *sb = ac->ac_sb;
1960 struct ext4_sb_info *sbi = EXT4_SB(sb);
1961 void *bitmap = e4b->bd_bitmap;
1962 struct ext4_free_extent ex;
1963 ext4_fsblk_t first_group_block;
1968 BUG_ON(sbi->s_stripe == 0);
1970 /* find first stripe-aligned block in group */
1971 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1973 a = first_group_block + sbi->s_stripe - 1;
1974 do_div(a, sbi->s_stripe);
1975 i = (a * sbi->s_stripe) - first_group_block;
1977 while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1978 if (!mb_test_bit(i, bitmap)) {
1979 max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1980 if (max >= sbi->s_stripe) {
1983 ext4_mb_use_best_found(ac, e4b);
1991 /* This is now called BEFORE we load the buddy bitmap. */
1992 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1993 ext4_group_t group, int cr)
1995 unsigned free, fragments;
1996 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1997 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1999 BUG_ON(cr < 0 || cr >= 4);
2001 free = grp->bb_free;
2004 if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2007 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2010 /* We only do this if the grp has never been initialized */
2011 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2012 int ret = ext4_mb_init_group(ac->ac_sb, group);
2017 fragments = grp->bb_fragments;
2023 BUG_ON(ac->ac_2order == 0);
2025 /* Avoid using the first bg of a flexgroup for data files */
2026 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2027 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2028 ((group % flex_size) == 0))
2031 if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2032 (free / fragments) >= ac->ac_g_ex.fe_len)
2035 if (grp->bb_largest_free_order < ac->ac_2order)
2040 if ((free / fragments) >= ac->ac_g_ex.fe_len)
2044 if (free >= ac->ac_g_ex.fe_len)
2056 static noinline_for_stack int
2057 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2059 ext4_group_t ngroups, group, i;
2062 struct ext4_sb_info *sbi;
2063 struct super_block *sb;
2064 struct ext4_buddy e4b;
2068 ngroups = ext4_get_groups_count(sb);
2069 /* non-extent files are limited to low blocks/groups */
2070 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2071 ngroups = sbi->s_blockfile_groups;
2073 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2075 /* first, try the goal */
2076 err = ext4_mb_find_by_goal(ac, &e4b);
2077 if (err || ac->ac_status == AC_STATUS_FOUND)
2080 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2084 * ac->ac2_order is set only if the fe_len is a power of 2
2085 * if ac2_order is set we also set criteria to 0 so that we
2086 * try exact allocation using buddy.
2088 i = fls(ac->ac_g_ex.fe_len);
2091 * We search using buddy data only if the order of the request
2092 * is greater than equal to the sbi_s_mb_order2_reqs
2093 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2095 if (i >= sbi->s_mb_order2_reqs) {
2097 * This should tell if fe_len is exactly power of 2
2099 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2100 ac->ac_2order = i - 1;
2103 /* if stream allocation is enabled, use global goal */
2104 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2105 /* TBD: may be hot point */
2106 spin_lock(&sbi->s_md_lock);
2107 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2108 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2109 spin_unlock(&sbi->s_md_lock);
2112 /* Let's just scan groups to find more-less suitable blocks */
2113 cr = ac->ac_2order ? 0 : 1;
2115 * cr == 0 try to get exact allocation,
2116 * cr == 3 try to get anything
2119 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2120 ac->ac_criteria = cr;
2122 * searching for the right group start
2123 * from the goal value specified
2125 group = ac->ac_g_ex.fe_group;
2127 for (i = 0; i < ngroups; group++, i++) {
2130 * Artificially restricted ngroups for non-extent
2131 * files makes group > ngroups possible on first loop.
2133 if (group >= ngroups)
2136 /* This now checks without needing the buddy page */
2137 if (!ext4_mb_good_group(ac, group, cr))
2140 err = ext4_mb_load_buddy(sb, group, &e4b);
2144 ext4_lock_group(sb, group);
2147 * We need to check again after locking the
2150 if (!ext4_mb_good_group(ac, group, cr)) {
2151 ext4_unlock_group(sb, group);
2152 ext4_mb_unload_buddy(&e4b);
2156 ac->ac_groups_scanned++;
2157 if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2158 ext4_mb_simple_scan_group(ac, &e4b);
2159 else if (cr == 1 && sbi->s_stripe &&
2160 !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2161 ext4_mb_scan_aligned(ac, &e4b);
2163 ext4_mb_complex_scan_group(ac, &e4b);
2165 ext4_unlock_group(sb, group);
2166 ext4_mb_unload_buddy(&e4b);
2168 if (ac->ac_status != AC_STATUS_CONTINUE)
2173 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2174 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2176 * We've been searching too long. Let's try to allocate
2177 * the best chunk we've found so far
2180 ext4_mb_try_best_found(ac, &e4b);
2181 if (ac->ac_status != AC_STATUS_FOUND) {
2183 * Someone more lucky has already allocated it.
2184 * The only thing we can do is just take first
2186 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2188 ac->ac_b_ex.fe_group = 0;
2189 ac->ac_b_ex.fe_start = 0;
2190 ac->ac_b_ex.fe_len = 0;
2191 ac->ac_status = AC_STATUS_CONTINUE;
2192 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2194 atomic_inc(&sbi->s_mb_lost_chunks);
2202 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2204 struct super_block *sb = seq->private;
2207 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2210 return (void *) ((unsigned long) group);
2213 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2215 struct super_block *sb = seq->private;
2219 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2222 return (void *) ((unsigned long) group);
2225 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2227 struct super_block *sb = seq->private;
2228 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2230 int err, buddy_loaded = 0;
2231 struct ext4_buddy e4b;
2232 struct ext4_group_info *grinfo;
2234 struct ext4_group_info info;
2235 ext4_grpblk_t counters[16];
2240 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2241 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2242 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2243 "group", "free", "frags", "first",
2244 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2245 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2247 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2248 sizeof(struct ext4_group_info);
2249 grinfo = ext4_get_group_info(sb, group);
2250 /* Load the group info in memory only if not already loaded. */
2251 if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2252 err = ext4_mb_load_buddy(sb, group, &e4b);
2254 seq_printf(seq, "#%-5u: I/O error\n", group);
2260 memcpy(&sg, ext4_get_group_info(sb, group), i);
2263 ext4_mb_unload_buddy(&e4b);
2265 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2266 sg.info.bb_fragments, sg.info.bb_first_free);
2267 for (i = 0; i <= 13; i++)
2268 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2269 sg.info.bb_counters[i] : 0);
2270 seq_printf(seq, " ]\n");
2275 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2279 static const struct seq_operations ext4_mb_seq_groups_ops = {
2280 .start = ext4_mb_seq_groups_start,
2281 .next = ext4_mb_seq_groups_next,
2282 .stop = ext4_mb_seq_groups_stop,
2283 .show = ext4_mb_seq_groups_show,
2286 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2288 struct super_block *sb = PDE_DATA(inode);
2291 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2293 struct seq_file *m = file->private_data;
2300 static const struct file_operations ext4_mb_seq_groups_fops = {
2301 .owner = THIS_MODULE,
2302 .open = ext4_mb_seq_groups_open,
2304 .llseek = seq_lseek,
2305 .release = seq_release,
2308 static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2310 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2311 struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2318 * Allocate the top-level s_group_info array for the specified number
2321 int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2323 struct ext4_sb_info *sbi = EXT4_SB(sb);
2325 struct ext4_group_info ***new_groupinfo;
2327 size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2328 EXT4_DESC_PER_BLOCK_BITS(sb);
2329 if (size <= sbi->s_group_info_size)
2332 size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2333 new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2334 if (!new_groupinfo) {
2335 ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2338 if (sbi->s_group_info) {
2339 memcpy(new_groupinfo, sbi->s_group_info,
2340 sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2341 ext4_kvfree(sbi->s_group_info);
2343 sbi->s_group_info = new_groupinfo;
2344 sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2345 ext4_debug("allocated s_groupinfo array for %d meta_bg's\n",
2346 sbi->s_group_info_size);
2350 /* Create and initialize ext4_group_info data for the given group. */
2351 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2352 struct ext4_group_desc *desc)
2356 struct ext4_sb_info *sbi = EXT4_SB(sb);
2357 struct ext4_group_info **meta_group_info;
2358 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2361 * First check if this group is the first of a reserved block.
2362 * If it's true, we have to allocate a new table of pointers
2363 * to ext4_group_info structures
2365 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2366 metalen = sizeof(*meta_group_info) <<
2367 EXT4_DESC_PER_BLOCK_BITS(sb);
2368 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2369 if (meta_group_info == NULL) {
2370 ext4_msg(sb, KERN_ERR, "can't allocate mem "
2371 "for a buddy group");
2372 goto exit_meta_group_info;
2374 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2379 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2380 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2382 meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2383 if (meta_group_info[i] == NULL) {
2384 ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2385 goto exit_group_info;
2387 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2388 &(meta_group_info[i]->bb_state));
2391 * initialize bb_free to be able to skip
2392 * empty groups without initialization
2394 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2395 meta_group_info[i]->bb_free =
2396 ext4_free_clusters_after_init(sb, group, desc);
2398 meta_group_info[i]->bb_free =
2399 ext4_free_group_clusters(sb, desc);
2402 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2403 init_rwsem(&meta_group_info[i]->alloc_sem);
2404 meta_group_info[i]->bb_free_root = RB_ROOT;
2405 meta_group_info[i]->bb_largest_free_order = -1; /* uninit */
2409 struct buffer_head *bh;
2410 meta_group_info[i]->bb_bitmap =
2411 kmalloc(sb->s_blocksize, GFP_KERNEL);
2412 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2413 bh = ext4_read_block_bitmap(sb, group);
2415 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2424 /* If a meta_group_info table has been allocated, release it now */
2425 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2426 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2427 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2429 exit_meta_group_info:
2431 } /* ext4_mb_add_groupinfo */
2433 static int ext4_mb_init_backend(struct super_block *sb)
2435 ext4_group_t ngroups = ext4_get_groups_count(sb);
2437 struct ext4_sb_info *sbi = EXT4_SB(sb);
2439 struct ext4_group_desc *desc;
2440 struct kmem_cache *cachep;
2442 err = ext4_mb_alloc_groupinfo(sb, ngroups);
2446 sbi->s_buddy_cache = new_inode(sb);
2447 if (sbi->s_buddy_cache == NULL) {
2448 ext4_msg(sb, KERN_ERR, "can't get new inode");
2451 /* To avoid potentially colliding with an valid on-disk inode number,
2452 * use EXT4_BAD_INO for the buddy cache inode number. This inode is
2453 * not in the inode hash, so it should never be found by iget(), but
2454 * this will avoid confusion if it ever shows up during debugging. */
2455 sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2456 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2457 for (i = 0; i < ngroups; i++) {
2458 desc = ext4_get_group_desc(sb, i, NULL);
2460 ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2463 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2470 cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2472 kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2473 i = sbi->s_group_info_size;
2475 kfree(sbi->s_group_info[i]);
2476 iput(sbi->s_buddy_cache);
2478 ext4_kvfree(sbi->s_group_info);
2482 static void ext4_groupinfo_destroy_slabs(void)
2486 for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2487 if (ext4_groupinfo_caches[i])
2488 kmem_cache_destroy(ext4_groupinfo_caches[i]);
2489 ext4_groupinfo_caches[i] = NULL;
2493 static int ext4_groupinfo_create_slab(size_t size)
2495 static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2497 int blocksize_bits = order_base_2(size);
2498 int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2499 struct kmem_cache *cachep;
2501 if (cache_index >= NR_GRPINFO_CACHES)
2504 if (unlikely(cache_index < 0))
2507 mutex_lock(&ext4_grpinfo_slab_create_mutex);
2508 if (ext4_groupinfo_caches[cache_index]) {
2509 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2510 return 0; /* Already created */
2513 slab_size = offsetof(struct ext4_group_info,
2514 bb_counters[blocksize_bits + 2]);
2516 cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2517 slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2520 ext4_groupinfo_caches[cache_index] = cachep;
2522 mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2525 "EXT4-fs: no memory for groupinfo slab cache\n");
2532 int ext4_mb_init(struct super_block *sb)
2534 struct ext4_sb_info *sbi = EXT4_SB(sb);
2540 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2542 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2543 if (sbi->s_mb_offsets == NULL) {
2548 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2549 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2550 if (sbi->s_mb_maxs == NULL) {
2555 ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2559 /* order 0 is regular bitmap */
2560 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2561 sbi->s_mb_offsets[0] = 0;
2565 max = sb->s_blocksize << 2;
2567 sbi->s_mb_offsets[i] = offset;
2568 sbi->s_mb_maxs[i] = max;
2569 offset += 1 << (sb->s_blocksize_bits - i);
2572 } while (i <= sb->s_blocksize_bits + 1);
2574 spin_lock_init(&sbi->s_md_lock);
2575 spin_lock_init(&sbi->s_bal_lock);
2577 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2578 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2579 sbi->s_mb_stats = MB_DEFAULT_STATS;
2580 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2581 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2583 * The default group preallocation is 512, which for 4k block
2584 * sizes translates to 2 megabytes. However for bigalloc file
2585 * systems, this is probably too big (i.e, if the cluster size
2586 * is 1 megabyte, then group preallocation size becomes half a
2587 * gigabyte!). As a default, we will keep a two megabyte
2588 * group pralloc size for cluster sizes up to 64k, and after
2589 * that, we will force a minimum group preallocation size of
2590 * 32 clusters. This translates to 8 megs when the cluster
2591 * size is 256k, and 32 megs when the cluster size is 1 meg,
2592 * which seems reasonable as a default.
2594 sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2595 sbi->s_cluster_bits, 32);
2597 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2598 * to the lowest multiple of s_stripe which is bigger than
2599 * the s_mb_group_prealloc as determined above. We want
2600 * the preallocation size to be an exact multiple of the
2601 * RAID stripe size so that preallocations don't fragment
2604 if (sbi->s_stripe > 1) {
2605 sbi->s_mb_group_prealloc = roundup(
2606 sbi->s_mb_group_prealloc, sbi->s_stripe);
2609 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2610 if (sbi->s_locality_groups == NULL) {
2612 goto out_free_groupinfo_slab;
2614 for_each_possible_cpu(i) {
2615 struct ext4_locality_group *lg;
2616 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2617 mutex_init(&lg->lg_mutex);
2618 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2619 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2620 spin_lock_init(&lg->lg_prealloc_lock);
2623 /* init file for buddy data */
2624 ret = ext4_mb_init_backend(sb);
2626 goto out_free_locality_groups;
2629 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2630 &ext4_mb_seq_groups_fops, sb);
2634 out_free_locality_groups:
2635 free_percpu(sbi->s_locality_groups);
2636 sbi->s_locality_groups = NULL;
2637 out_free_groupinfo_slab:
2638 ext4_groupinfo_destroy_slabs();
2640 kfree(sbi->s_mb_offsets);
2641 sbi->s_mb_offsets = NULL;
2642 kfree(sbi->s_mb_maxs);
2643 sbi->s_mb_maxs = NULL;
2647 /* need to called with the ext4 group lock held */
2648 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2650 struct ext4_prealloc_space *pa;
2651 struct list_head *cur, *tmp;
2654 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2655 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2656 list_del(&pa->pa_group_list);
2658 kmem_cache_free(ext4_pspace_cachep, pa);
2661 mb_debug(1, "mballoc: %u PAs left\n", count);
2665 int ext4_mb_release(struct super_block *sb)
2667 ext4_group_t ngroups = ext4_get_groups_count(sb);
2669 int num_meta_group_infos;
2670 struct ext4_group_info *grinfo;
2671 struct ext4_sb_info *sbi = EXT4_SB(sb);
2672 struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2675 remove_proc_entry("mb_groups", sbi->s_proc);
2677 if (sbi->s_group_info) {
2678 for (i = 0; i < ngroups; i++) {
2679 grinfo = ext4_get_group_info(sb, i);
2681 kfree(grinfo->bb_bitmap);
2683 ext4_lock_group(sb, i);
2684 ext4_mb_cleanup_pa(grinfo);
2685 ext4_unlock_group(sb, i);
2686 kmem_cache_free(cachep, grinfo);
2688 num_meta_group_infos = (ngroups +
2689 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2690 EXT4_DESC_PER_BLOCK_BITS(sb);
2691 for (i = 0; i < num_meta_group_infos; i++)
2692 kfree(sbi->s_group_info[i]);
2693 ext4_kvfree(sbi->s_group_info);
2695 kfree(sbi->s_mb_offsets);
2696 kfree(sbi->s_mb_maxs);
2697 if (sbi->s_buddy_cache)
2698 iput(sbi->s_buddy_cache);
2699 if (sbi->s_mb_stats) {
2700 ext4_msg(sb, KERN_INFO,
2701 "mballoc: %u blocks %u reqs (%u success)",
2702 atomic_read(&sbi->s_bal_allocated),
2703 atomic_read(&sbi->s_bal_reqs),
2704 atomic_read(&sbi->s_bal_success));
2705 ext4_msg(sb, KERN_INFO,
2706 "mballoc: %u extents scanned, %u goal hits, "
2707 "%u 2^N hits, %u breaks, %u lost",
2708 atomic_read(&sbi->s_bal_ex_scanned),
2709 atomic_read(&sbi->s_bal_goals),
2710 atomic_read(&sbi->s_bal_2orders),
2711 atomic_read(&sbi->s_bal_breaks),
2712 atomic_read(&sbi->s_mb_lost_chunks));
2713 ext4_msg(sb, KERN_INFO,
2714 "mballoc: %lu generated and it took %Lu",
2715 sbi->s_mb_buddies_generated,
2716 sbi->s_mb_generation_time);
2717 ext4_msg(sb, KERN_INFO,
2718 "mballoc: %u preallocated, %u discarded",
2719 atomic_read(&sbi->s_mb_preallocated),
2720 atomic_read(&sbi->s_mb_discarded));
2723 free_percpu(sbi->s_locality_groups);
2728 static inline int ext4_issue_discard(struct super_block *sb,
2729 ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2731 ext4_fsblk_t discard_block;
2733 discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2734 ext4_group_first_block_no(sb, block_group));
2735 count = EXT4_C2B(EXT4_SB(sb), count);
2736 trace_ext4_discard_blocks(sb,
2737 (unsigned long long) discard_block, count);
2738 return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2742 * This function is called by the jbd2 layer once the commit has finished,
2743 * so we know we can free the blocks that were released with that commit.
2745 static void ext4_free_data_callback(struct super_block *sb,
2746 struct ext4_journal_cb_entry *jce,
2749 struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2750 struct ext4_buddy e4b;
2751 struct ext4_group_info *db;
2752 int err, count = 0, count2 = 0;
2754 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2755 entry->efd_count, entry->efd_group, entry);
2757 if (test_opt(sb, DISCARD)) {
2758 err = ext4_issue_discard(sb, entry->efd_group,
2759 entry->efd_start_cluster,
2761 if (err && err != -EOPNOTSUPP)
2762 ext4_msg(sb, KERN_WARNING, "discard request in"
2763 " group:%d block:%d count:%d failed"
2764 " with %d", entry->efd_group,
2765 entry->efd_start_cluster,
2766 entry->efd_count, err);
2769 err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2770 /* we expect to find existing buddy because it's pinned */
2775 /* there are blocks to put in buddy to make them really free */
2776 count += entry->efd_count;
2778 ext4_lock_group(sb, entry->efd_group);
2779 /* Take it out of per group rb tree */
2780 rb_erase(&entry->efd_node, &(db->bb_free_root));
2781 mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2784 * Clear the trimmed flag for the group so that the next
2785 * ext4_trim_fs can trim it.
2786 * If the volume is mounted with -o discard, online discard
2787 * is supported and the free blocks will be trimmed online.
2789 if (!test_opt(sb, DISCARD))
2790 EXT4_MB_GRP_CLEAR_TRIMMED(db);
2792 if (!db->bb_free_root.rb_node) {
2793 /* No more items in the per group rb tree
2794 * balance refcounts from ext4_mb_free_metadata()
2796 page_cache_release(e4b.bd_buddy_page);
2797 page_cache_release(e4b.bd_bitmap_page);
2799 ext4_unlock_group(sb, entry->efd_group);
2800 kmem_cache_free(ext4_free_data_cachep, entry);
2801 ext4_mb_unload_buddy(&e4b);
2803 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2806 int __init ext4_init_mballoc(void)
2808 ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2809 SLAB_RECLAIM_ACCOUNT);
2810 if (ext4_pspace_cachep == NULL)
2813 ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2814 SLAB_RECLAIM_ACCOUNT);
2815 if (ext4_ac_cachep == NULL) {
2816 kmem_cache_destroy(ext4_pspace_cachep);
2820 ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2821 SLAB_RECLAIM_ACCOUNT);
2822 if (ext4_free_data_cachep == NULL) {
2823 kmem_cache_destroy(ext4_pspace_cachep);
2824 kmem_cache_destroy(ext4_ac_cachep);
2830 void ext4_exit_mballoc(void)
2833 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2834 * before destroying the slab cache.
2837 kmem_cache_destroy(ext4_pspace_cachep);
2838 kmem_cache_destroy(ext4_ac_cachep);
2839 kmem_cache_destroy(ext4_free_data_cachep);
2840 ext4_groupinfo_destroy_slabs();
2845 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2846 * Returns 0 if success or error code
2848 static noinline_for_stack int
2849 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2850 handle_t *handle, unsigned int reserv_clstrs)
2852 struct buffer_head *bitmap_bh = NULL;
2853 struct ext4_group_desc *gdp;
2854 struct buffer_head *gdp_bh;
2855 struct ext4_sb_info *sbi;
2856 struct super_block *sb;
2860 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2861 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2867 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2871 err = ext4_journal_get_write_access(handle, bitmap_bh);
2876 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2880 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2881 ext4_free_group_clusters(sb, gdp));
2883 err = ext4_journal_get_write_access(handle, gdp_bh);
2887 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2889 len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2890 if (!ext4_data_block_valid(sbi, block, len)) {
2891 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2892 "fs metadata", block, block+len);
2893 /* File system mounted not to panic on error
2894 * Fix the bitmap and repeat the block allocation
2895 * We leak some of the blocks here.
2897 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2898 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2899 ac->ac_b_ex.fe_len);
2900 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2901 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2907 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2908 #ifdef AGGRESSIVE_CHECK
2911 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2912 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2913 bitmap_bh->b_data));
2917 ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2918 ac->ac_b_ex.fe_len);
2919 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2920 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2921 ext4_free_group_clusters_set(sb, gdp,
2922 ext4_free_clusters_after_init(sb,
2923 ac->ac_b_ex.fe_group, gdp));
2925 len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2926 ext4_free_group_clusters_set(sb, gdp, len);
2927 ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2928 ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2930 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2931 percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2933 * Now reduce the dirty block count also. Should not go negative
2935 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2936 /* release all the reserved blocks if non delalloc */
2937 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2940 if (sbi->s_log_groups_per_flex) {
2941 ext4_group_t flex_group = ext4_flex_group(sbi,
2942 ac->ac_b_ex.fe_group);
2943 atomic64_sub(ac->ac_b_ex.fe_len,
2944 &sbi->s_flex_groups[flex_group].free_clusters);
2947 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2950 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2958 * here we normalize request for locality group
2959 * Group request are normalized to s_mb_group_prealloc, which goes to
2960 * s_strip if we set the same via mount option.
2961 * s_mb_group_prealloc can be configured via
2962 * /sys/fs/ext4/<partition>/mb_group_prealloc
2964 * XXX: should we try to preallocate more than the group has now?
2966 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2968 struct super_block *sb = ac->ac_sb;
2969 struct ext4_locality_group *lg = ac->ac_lg;
2972 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2973 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2974 current->pid, ac->ac_g_ex.fe_len);
2978 * Normalization means making request better in terms of
2979 * size and alignment
2981 static noinline_for_stack void
2982 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2983 struct ext4_allocation_request *ar)
2985 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2988 loff_t size, start_off;
2989 loff_t orig_size __maybe_unused;
2991 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2992 struct ext4_prealloc_space *pa;
2994 /* do normalize only data requests, metadata requests
2995 do not need preallocation */
2996 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2999 /* sometime caller may want exact blocks */
3000 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3003 /* caller may indicate that preallocation isn't
3004 * required (it's a tail, for example) */
3005 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3008 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3009 ext4_mb_normalize_group_request(ac);
3013 bsbits = ac->ac_sb->s_blocksize_bits;
3015 /* first, let's learn actual file size
3016 * given current request is allocated */
3017 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3018 size = size << bsbits;
3019 if (size < i_size_read(ac->ac_inode))
3020 size = i_size_read(ac->ac_inode);
3023 /* max size of free chunks */
3026 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3027 (req <= (size) || max <= (chunk_size))
3029 /* first, try to predict filesize */
3030 /* XXX: should this table be tunable? */
3032 if (size <= 16 * 1024) {
3034 } else if (size <= 32 * 1024) {
3036 } else if (size <= 64 * 1024) {
3038 } else if (size <= 128 * 1024) {
3040 } else if (size <= 256 * 1024) {
3042 } else if (size <= 512 * 1024) {
3044 } else if (size <= 1024 * 1024) {
3046 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3047 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3048 (21 - bsbits)) << 21;
3049 size = 2 * 1024 * 1024;
3050 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3051 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3052 (22 - bsbits)) << 22;
3053 size = 4 * 1024 * 1024;
3054 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3055 (8<<20)>>bsbits, max, 8 * 1024)) {
3056 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3057 (23 - bsbits)) << 23;
3058 size = 8 * 1024 * 1024;
3060 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
3061 size = ac->ac_o_ex.fe_len << bsbits;
3063 size = size >> bsbits;
3064 start = start_off >> bsbits;
3066 /* don't cover already allocated blocks in selected range */
3067 if (ar->pleft && start <= ar->lleft) {
3068 size -= ar->lleft + 1 - start;
3069 start = ar->lleft + 1;
3071 if (ar->pright && start + size - 1 >= ar->lright)
3072 size -= start + size - ar->lright;
3076 /* check we don't cross already preallocated blocks */
3078 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3083 spin_lock(&pa->pa_lock);
3084 if (pa->pa_deleted) {
3085 spin_unlock(&pa->pa_lock);
3089 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3092 /* PA must not overlap original request */
3093 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3094 ac->ac_o_ex.fe_logical < pa->pa_lstart));
3096 /* skip PAs this normalized request doesn't overlap with */
3097 if (pa->pa_lstart >= end || pa_end <= start) {
3098 spin_unlock(&pa->pa_lock);
3101 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3103 /* adjust start or end to be adjacent to this pa */
3104 if (pa_end <= ac->ac_o_ex.fe_logical) {
3105 BUG_ON(pa_end < start);
3107 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3108 BUG_ON(pa->pa_lstart > end);
3109 end = pa->pa_lstart;
3111 spin_unlock(&pa->pa_lock);
3116 /* XXX: extra loop to check we really don't overlap preallocations */
3118 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3121 spin_lock(&pa->pa_lock);
3122 if (pa->pa_deleted == 0) {
3123 pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3125 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3127 spin_unlock(&pa->pa_lock);
3131 if (start + size <= ac->ac_o_ex.fe_logical &&
3132 start > ac->ac_o_ex.fe_logical) {
3133 ext4_msg(ac->ac_sb, KERN_ERR,
3134 "start %lu, size %lu, fe_logical %lu",
3135 (unsigned long) start, (unsigned long) size,
3136 (unsigned long) ac->ac_o_ex.fe_logical);
3138 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3139 start > ac->ac_o_ex.fe_logical);
3140 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3142 /* now prepare goal request */
3144 /* XXX: is it better to align blocks WRT to logical
3145 * placement or satisfy big request as is */
3146 ac->ac_g_ex.fe_logical = start;
3147 ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3149 /* define goal start in order to merge */
3150 if (ar->pright && (ar->lright == (start + size))) {
3151 /* merge to the right */
3152 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3153 &ac->ac_f_ex.fe_group,
3154 &ac->ac_f_ex.fe_start);
3155 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3157 if (ar->pleft && (ar->lleft + 1 == start)) {
3158 /* merge to the left */
3159 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3160 &ac->ac_f_ex.fe_group,
3161 &ac->ac_f_ex.fe_start);
3162 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3165 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3166 (unsigned) orig_size, (unsigned) start);
3169 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3171 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3173 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3174 atomic_inc(&sbi->s_bal_reqs);
3175 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3176 if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3177 atomic_inc(&sbi->s_bal_success);
3178 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3179 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3180 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3181 atomic_inc(&sbi->s_bal_goals);
3182 if (ac->ac_found > sbi->s_mb_max_to_scan)
3183 atomic_inc(&sbi->s_bal_breaks);
3186 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3187 trace_ext4_mballoc_alloc(ac);
3189 trace_ext4_mballoc_prealloc(ac);
3193 * Called on failure; free up any blocks from the inode PA for this
3194 * context. We don't need this for MB_GROUP_PA because we only change
3195 * pa_free in ext4_mb_release_context(), but on failure, we've already
3196 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3198 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3200 struct ext4_prealloc_space *pa = ac->ac_pa;
3201 struct ext4_buddy e4b;
3205 if (ac->ac_f_ex.fe_len == 0)
3207 err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3210 * This should never happen since we pin the
3211 * pages in the ext4_allocation_context so
3212 * ext4_mb_load_buddy() should never fail.
3214 WARN(1, "mb_load_buddy failed (%d)", err);
3217 ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3218 mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3219 ac->ac_f_ex.fe_len);
3220 ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3221 ext4_mb_unload_buddy(&e4b);
3224 if (pa->pa_type == MB_INODE_PA)
3225 pa->pa_free += ac->ac_b_ex.fe_len;
3229 * use blocks preallocated to inode
3231 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3232 struct ext4_prealloc_space *pa)
3234 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3239 /* found preallocated blocks, use them */
3240 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3241 end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3242 start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3243 len = EXT4_NUM_B2C(sbi, end - start);
3244 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3245 &ac->ac_b_ex.fe_start);
3246 ac->ac_b_ex.fe_len = len;
3247 ac->ac_status = AC_STATUS_FOUND;
3250 BUG_ON(start < pa->pa_pstart);
3251 BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3252 BUG_ON(pa->pa_free < len);
3255 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3259 * use blocks preallocated to locality group
3261 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3262 struct ext4_prealloc_space *pa)
3264 unsigned int len = ac->ac_o_ex.fe_len;
3266 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3267 &ac->ac_b_ex.fe_group,
3268 &ac->ac_b_ex.fe_start);
3269 ac->ac_b_ex.fe_len = len;
3270 ac->ac_status = AC_STATUS_FOUND;
3273 /* we don't correct pa_pstart or pa_plen here to avoid
3274 * possible race when the group is being loaded concurrently
3275 * instead we correct pa later, after blocks are marked
3276 * in on-disk bitmap -- see ext4_mb_release_context()
3277 * Other CPUs are prevented from allocating from this pa by lg_mutex
3279 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3283 * Return the prealloc space that have minimal distance
3284 * from the goal block. @cpa is the prealloc
3285 * space that is having currently known minimal distance
3286 * from the goal block.
3288 static struct ext4_prealloc_space *
3289 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3290 struct ext4_prealloc_space *pa,
3291 struct ext4_prealloc_space *cpa)
3293 ext4_fsblk_t cur_distance, new_distance;
3296 atomic_inc(&pa->pa_count);
3299 cur_distance = abs(goal_block - cpa->pa_pstart);
3300 new_distance = abs(goal_block - pa->pa_pstart);
3302 if (cur_distance <= new_distance)
3305 /* drop the previous reference */
3306 atomic_dec(&cpa->pa_count);
3307 atomic_inc(&pa->pa_count);
3312 * search goal blocks in preallocated space
3314 static noinline_for_stack int
3315 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3317 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3319 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3320 struct ext4_locality_group *lg;
3321 struct ext4_prealloc_space *pa, *cpa = NULL;
3322 ext4_fsblk_t goal_block;
3324 /* only data can be preallocated */
3325 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3328 /* first, try per-file preallocation */
3330 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3332 /* all fields in this condition don't change,
3333 * so we can skip locking for them */
3334 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3335 ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3336 EXT4_C2B(sbi, pa->pa_len)))
3339 /* non-extent files can't have physical blocks past 2^32 */
3340 if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3341 (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3342 EXT4_MAX_BLOCK_FILE_PHYS))
3345 /* found preallocated blocks, use them */
3346 spin_lock(&pa->pa_lock);
3347 if (pa->pa_deleted == 0 && pa->pa_free) {
3348 atomic_inc(&pa->pa_count);
3349 ext4_mb_use_inode_pa(ac, pa);
3350 spin_unlock(&pa->pa_lock);
3351 ac->ac_criteria = 10;
3355 spin_unlock(&pa->pa_lock);
3359 /* can we use group allocation? */
3360 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3363 /* inode may have no locality group for some reason */
3367 order = fls(ac->ac_o_ex.fe_len) - 1;
3368 if (order > PREALLOC_TB_SIZE - 1)
3369 /* The max size of hash table is PREALLOC_TB_SIZE */
3370 order = PREALLOC_TB_SIZE - 1;
3372 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3374 * search for the prealloc space that is having
3375 * minimal distance from the goal block.
3377 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3379 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3381 spin_lock(&pa->pa_lock);
3382 if (pa->pa_deleted == 0 &&
3383 pa->pa_free >= ac->ac_o_ex.fe_len) {
3385 cpa = ext4_mb_check_group_pa(goal_block,
3388 spin_unlock(&pa->pa_lock);
3393 ext4_mb_use_group_pa(ac, cpa);
3394 ac->ac_criteria = 20;
3401 * the function goes through all block freed in the group
3402 * but not yet committed and marks them used in in-core bitmap.
3403 * buddy must be generated from this bitmap
3404 * Need to be called with the ext4 group lock held
3406 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3410 struct ext4_group_info *grp;
3411 struct ext4_free_data *entry;
3413 grp = ext4_get_group_info(sb, group);
3414 n = rb_first(&(grp->bb_free_root));
3417 entry = rb_entry(n, struct ext4_free_data, efd_node);
3418 ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3425 * the function goes through all preallocation in this group and marks them
3426 * used in in-core bitmap. buddy must be generated from this bitmap
3427 * Need to be called with ext4 group lock held
3429 static noinline_for_stack
3430 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3433 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3434 struct ext4_prealloc_space *pa;
3435 struct list_head *cur;
3436 ext4_group_t groupnr;
3437 ext4_grpblk_t start;
3438 int preallocated = 0;
3441 /* all form of preallocation discards first load group,
3442 * so the only competing code is preallocation use.
3443 * we don't need any locking here
3444 * notice we do NOT ignore preallocations with pa_deleted
3445 * otherwise we could leave used blocks available for
3446 * allocation in buddy when concurrent ext4_mb_put_pa()
3447 * is dropping preallocation
3449 list_for_each(cur, &grp->bb_prealloc_list) {
3450 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3451 spin_lock(&pa->pa_lock);
3452 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3455 spin_unlock(&pa->pa_lock);
3456 if (unlikely(len == 0))
3458 BUG_ON(groupnr != group);
3459 ext4_set_bits(bitmap, start, len);
3460 preallocated += len;
3462 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3465 static void ext4_mb_pa_callback(struct rcu_head *head)
3467 struct ext4_prealloc_space *pa;
3468 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3470 BUG_ON(atomic_read(&pa->pa_count));
3471 BUG_ON(pa->pa_deleted == 0);
3472 kmem_cache_free(ext4_pspace_cachep, pa);
3476 * drops a reference to preallocated space descriptor
3477 * if this was the last reference and the space is consumed
3479 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3480 struct super_block *sb, struct ext4_prealloc_space *pa)
3483 ext4_fsblk_t grp_blk;
3485 /* in this short window concurrent discard can set pa_deleted */
3486 spin_lock(&pa->pa_lock);
3487 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3488 spin_unlock(&pa->pa_lock);
3492 if (pa->pa_deleted == 1) {
3493 spin_unlock(&pa->pa_lock);
3498 spin_unlock(&pa->pa_lock);
3500 grp_blk = pa->pa_pstart;
3502 * If doing group-based preallocation, pa_pstart may be in the
3503 * next group when pa is used up
3505 if (pa->pa_type == MB_GROUP_PA)
3508 grp = ext4_get_group_number(sb, grp_blk);
3513 * P1 (buddy init) P2 (regular allocation)
3514 * find block B in PA
3515 * copy on-disk bitmap to buddy
3516 * mark B in on-disk bitmap
3517 * drop PA from group
3518 * mark all PAs in buddy
3520 * thus, P1 initializes buddy with B available. to prevent this
3521 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3524 ext4_lock_group(sb, grp);
3525 list_del(&pa->pa_group_list);
3526 ext4_unlock_group(sb, grp);
3528 spin_lock(pa->pa_obj_lock);
3529 list_del_rcu(&pa->pa_inode_list);
3530 spin_unlock(pa->pa_obj_lock);
3532 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3536 * creates new preallocated space for given inode
3538 static noinline_for_stack int
3539 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3541 struct super_block *sb = ac->ac_sb;
3542 struct ext4_sb_info *sbi = EXT4_SB(sb);
3543 struct ext4_prealloc_space *pa;
3544 struct ext4_group_info *grp;
3545 struct ext4_inode_info *ei;
3547 /* preallocate only when found space is larger then requested */
3548 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3549 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3550 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3552 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3556 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3562 /* we can't allocate as much as normalizer wants.
3563 * so, found space must get proper lstart
3564 * to cover original request */
3565 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3566 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3568 /* we're limited by original request in that
3569 * logical block must be covered any way
3570 * winl is window we can move our chunk within */
3571 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3573 /* also, we should cover whole original request */
3574 wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3576 /* the smallest one defines real window */
3577 win = min(winl, wins);
3579 offs = ac->ac_o_ex.fe_logical %
3580 EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3581 if (offs && offs < win)
3584 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3585 EXT4_NUM_B2C(sbi, win);
3586 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3587 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3590 /* preallocation can change ac_b_ex, thus we store actually
3591 * allocated blocks for history */
3592 ac->ac_f_ex = ac->ac_b_ex;
3594 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3595 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3596 pa->pa_len = ac->ac_b_ex.fe_len;
3597 pa->pa_free = pa->pa_len;
3598 atomic_set(&pa->pa_count, 1);
3599 spin_lock_init(&pa->pa_lock);
3600 INIT_LIST_HEAD(&pa->pa_inode_list);
3601 INIT_LIST_HEAD(&pa->pa_group_list);
3603 pa->pa_type = MB_INODE_PA;
3605 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3606 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3607 trace_ext4_mb_new_inode_pa(ac, pa);
3609 ext4_mb_use_inode_pa(ac, pa);
3610 atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3612 ei = EXT4_I(ac->ac_inode);
3613 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3615 pa->pa_obj_lock = &ei->i_prealloc_lock;
3616 pa->pa_inode = ac->ac_inode;
3618 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3619 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3620 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3622 spin_lock(pa->pa_obj_lock);
3623 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3624 spin_unlock(pa->pa_obj_lock);
3630 * creates new preallocated space for locality group inodes belongs to
3632 static noinline_for_stack int
3633 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3635 struct super_block *sb = ac->ac_sb;
3636 struct ext4_locality_group *lg;
3637 struct ext4_prealloc_space *pa;
3638 struct ext4_group_info *grp;
3640 /* preallocate only when found space is larger then requested */
3641 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3642 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3643 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3645 BUG_ON(ext4_pspace_cachep == NULL);
3646 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3650 /* preallocation can change ac_b_ex, thus we store actually
3651 * allocated blocks for history */
3652 ac->ac_f_ex = ac->ac_b_ex;
3654 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3655 pa->pa_lstart = pa->pa_pstart;
3656 pa->pa_len = ac->ac_b_ex.fe_len;
3657 pa->pa_free = pa->pa_len;
3658 atomic_set(&pa->pa_count, 1);
3659 spin_lock_init(&pa->pa_lock);
3660 INIT_LIST_HEAD(&pa->pa_inode_list);
3661 INIT_LIST_HEAD(&pa->pa_group_list);
3663 pa->pa_type = MB_GROUP_PA;
3665 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3666 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3667 trace_ext4_mb_new_group_pa(ac, pa);
3669 ext4_mb_use_group_pa(ac, pa);
3670 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3672 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3676 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3677 pa->pa_inode = NULL;
3679 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3680 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3681 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3684 * We will later add the new pa to the right bucket
3685 * after updating the pa_free in ext4_mb_release_context
3690 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3694 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3695 err = ext4_mb_new_group_pa(ac);
3697 err = ext4_mb_new_inode_pa(ac);
3702 * finds all unused blocks in on-disk bitmap, frees them in
3703 * in-core bitmap and buddy.
3704 * @pa must be unlinked from inode and group lists, so that
3705 * nobody else can find/use it.
3706 * the caller MUST hold group/inode locks.
3707 * TODO: optimize the case when there are no in-core structures yet
3709 static noinline_for_stack int
3710 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3711 struct ext4_prealloc_space *pa)
3713 struct super_block *sb = e4b->bd_sb;
3714 struct ext4_sb_info *sbi = EXT4_SB(sb);
3719 unsigned long long grp_blk_start;
3723 BUG_ON(pa->pa_deleted == 0);
3724 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3725 grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3726 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3727 end = bit + pa->pa_len;
3730 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3733 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3734 mb_debug(1, " free preallocated %u/%u in group %u\n",
3735 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3736 (unsigned) next - bit, (unsigned) group);
3739 trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3740 trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3741 EXT4_C2B(sbi, bit)),
3743 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3746 if (free != pa->pa_free) {
3747 ext4_msg(e4b->bd_sb, KERN_CRIT,
3748 "pa %p: logic %lu, phys. %lu, len %lu",
3749 pa, (unsigned long) pa->pa_lstart,
3750 (unsigned long) pa->pa_pstart,
3751 (unsigned long) pa->pa_len);
3752 ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3755 * pa is already deleted so we use the value obtained
3756 * from the bitmap and continue.
3759 atomic_add(free, &sbi->s_mb_discarded);
3764 static noinline_for_stack int
3765 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3766 struct ext4_prealloc_space *pa)
3768 struct super_block *sb = e4b->bd_sb;
3772 trace_ext4_mb_release_group_pa(sb, pa);
3773 BUG_ON(pa->pa_deleted == 0);
3774 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3775 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3776 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3777 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3778 trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3784 * releases all preallocations in given group
3786 * first, we need to decide discard policy:
3787 * - when do we discard
3789 * - how many do we discard
3790 * 1) how many requested
3792 static noinline_for_stack int
3793 ext4_mb_discard_group_preallocations(struct super_block *sb,
3794 ext4_group_t group, int needed)
3796 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3797 struct buffer_head *bitmap_bh = NULL;
3798 struct ext4_prealloc_space *pa, *tmp;
3799 struct list_head list;
3800 struct ext4_buddy e4b;
3805 mb_debug(1, "discard preallocation for group %u\n", group);
3807 if (list_empty(&grp->bb_prealloc_list))
3810 bitmap_bh = ext4_read_block_bitmap(sb, group);
3811 if (bitmap_bh == NULL) {
3812 ext4_error(sb, "Error reading block bitmap for %u", group);
3816 err = ext4_mb_load_buddy(sb, group, &e4b);
3818 ext4_error(sb, "Error loading buddy information for %u", group);
3824 needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3826 INIT_LIST_HEAD(&list);
3828 ext4_lock_group(sb, group);
3829 list_for_each_entry_safe(pa, tmp,
3830 &grp->bb_prealloc_list, pa_group_list) {
3831 spin_lock(&pa->pa_lock);
3832 if (atomic_read(&pa->pa_count)) {
3833 spin_unlock(&pa->pa_lock);
3837 if (pa->pa_deleted) {
3838 spin_unlock(&pa->pa_lock);
3842 /* seems this one can be freed ... */
3845 /* we can trust pa_free ... */
3846 free += pa->pa_free;
3848 spin_unlock(&pa->pa_lock);
3850 list_del(&pa->pa_group_list);
3851 list_add(&pa->u.pa_tmp_list, &list);
3854 /* if we still need more blocks and some PAs were used, try again */
3855 if (free < needed && busy) {
3857 ext4_unlock_group(sb, group);
3862 /* found anything to free? */
3863 if (list_empty(&list)) {
3868 /* now free all selected PAs */
3869 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3871 /* remove from object (inode or locality group) */
3872 spin_lock(pa->pa_obj_lock);
3873 list_del_rcu(&pa->pa_inode_list);
3874 spin_unlock(pa->pa_obj_lock);
3876 if (pa->pa_type == MB_GROUP_PA)
3877 ext4_mb_release_group_pa(&e4b, pa);
3879 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3881 list_del(&pa->u.pa_tmp_list);
3882 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3886 ext4_unlock_group(sb, group);
3887 ext4_mb_unload_buddy(&e4b);
3893 * releases all non-used preallocated blocks for given inode
3895 * It's important to discard preallocations under i_data_sem
3896 * We don't want another block to be served from the prealloc
3897 * space when we are discarding the inode prealloc space.
3899 * FIXME!! Make sure it is valid at all the call sites
3901 void ext4_discard_preallocations(struct inode *inode)
3903 struct ext4_inode_info *ei = EXT4_I(inode);
3904 struct super_block *sb = inode->i_sb;
3905 struct buffer_head *bitmap_bh = NULL;
3906 struct ext4_prealloc_space *pa, *tmp;
3907 ext4_group_t group = 0;
3908 struct list_head list;
3909 struct ext4_buddy e4b;
3912 if (!S_ISREG(inode->i_mode)) {
3913 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3917 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3918 trace_ext4_discard_preallocations(inode);
3920 INIT_LIST_HEAD(&list);
3923 /* first, collect all pa's in the inode */
3924 spin_lock(&ei->i_prealloc_lock);
3925 while (!list_empty(&ei->i_prealloc_list)) {
3926 pa = list_entry(ei->i_prealloc_list.next,
3927 struct ext4_prealloc_space, pa_inode_list);
3928 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3929 spin_lock(&pa->pa_lock);
3930 if (atomic_read(&pa->pa_count)) {
3931 /* this shouldn't happen often - nobody should
3932 * use preallocation while we're discarding it */
3933 spin_unlock(&pa->pa_lock);
3934 spin_unlock(&ei->i_prealloc_lock);
3935 ext4_msg(sb, KERN_ERR,
3936 "uh-oh! used pa while discarding");
3938 schedule_timeout_uninterruptible(HZ);
3942 if (pa->pa_deleted == 0) {
3944 spin_unlock(&pa->pa_lock);
3945 list_del_rcu(&pa->pa_inode_list);
3946 list_add(&pa->u.pa_tmp_list, &list);
3950 /* someone is deleting pa right now */
3951 spin_unlock(&pa->pa_lock);
3952 spin_unlock(&ei->i_prealloc_lock);
3954 /* we have to wait here because pa_deleted
3955 * doesn't mean pa is already unlinked from
3956 * the list. as we might be called from
3957 * ->clear_inode() the inode will get freed
3958 * and concurrent thread which is unlinking
3959 * pa from inode's list may access already
3960 * freed memory, bad-bad-bad */
3962 /* XXX: if this happens too often, we can
3963 * add a flag to force wait only in case
3964 * of ->clear_inode(), but not in case of
3965 * regular truncate */
3966 schedule_timeout_uninterruptible(HZ);
3969 spin_unlock(&ei->i_prealloc_lock);
3971 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3972 BUG_ON(pa->pa_type != MB_INODE_PA);
3973 group = ext4_get_group_number(sb, pa->pa_pstart);
3975 err = ext4_mb_load_buddy(sb, group, &e4b);
3977 ext4_error(sb, "Error loading buddy information for %u",
3982 bitmap_bh = ext4_read_block_bitmap(sb, group);
3983 if (bitmap_bh == NULL) {
3984 ext4_error(sb, "Error reading block bitmap for %u",
3986 ext4_mb_unload_buddy(&e4b);
3990 ext4_lock_group(sb, group);
3991 list_del(&pa->pa_group_list);
3992 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3993 ext4_unlock_group(sb, group);
3995 ext4_mb_unload_buddy(&e4b);
3998 list_del(&pa->u.pa_tmp_list);
3999 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4003 #ifdef CONFIG_EXT4_DEBUG
4004 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4006 struct super_block *sb = ac->ac_sb;
4007 ext4_group_t ngroups, i;
4009 if (!ext4_mballoc_debug ||
4010 (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4013 ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4014 " Allocation context details:");
4015 ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4016 ac->ac_status, ac->ac_flags);
4017 ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4018 "goal %lu/%lu/%lu@%lu, "
4019 "best %lu/%lu/%lu@%lu cr %d",
4020 (unsigned long)ac->ac_o_ex.fe_group,
4021 (unsigned long)ac->ac_o_ex.fe_start,
4022 (unsigned long)ac->ac_o_ex.fe_len,
4023 (unsigned long)ac->ac_o_ex.fe_logical,
4024 (unsigned long)ac->ac_g_ex.fe_group,
4025 (unsigned long)ac->ac_g_ex.fe_start,
4026 (unsigned long)ac->ac_g_ex.fe_len,
4027 (unsigned long)ac->ac_g_ex.fe_logical,
4028 (unsigned long)ac->ac_b_ex.fe_group,
4029 (unsigned long)ac->ac_b_ex.fe_start,
4030 (unsigned long)ac->ac_b_ex.fe_len,
4031 (unsigned long)ac->ac_b_ex.fe_logical,
4032 (int)ac->ac_criteria);
4033 ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
4034 ac->ac_ex_scanned, ac->ac_found);
4035 ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4036 ngroups = ext4_get_groups_count(sb);
4037 for (i = 0; i < ngroups; i++) {
4038 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4039 struct ext4_prealloc_space *pa;
4040 ext4_grpblk_t start;
4041 struct list_head *cur;
4042 ext4_lock_group(sb, i);
4043 list_for_each(cur, &grp->bb_prealloc_list) {
4044 pa = list_entry(cur, struct ext4_prealloc_space,
4046 spin_lock(&pa->pa_lock);
4047 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4049 spin_unlock(&pa->pa_lock);
4050 printk(KERN_ERR "PA:%u:%d:%u \n", i,
4053 ext4_unlock_group(sb, i);
4055 if (grp->bb_free == 0)
4057 printk(KERN_ERR "%u: %d/%d \n",
4058 i, grp->bb_free, grp->bb_fragments);
4060 printk(KERN_ERR "\n");
4063 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4070 * We use locality group preallocation for small size file. The size of the
4071 * file is determined by the current size or the resulting size after
4072 * allocation which ever is larger
4074 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4076 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4078 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4079 int bsbits = ac->ac_sb->s_blocksize_bits;
4082 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4085 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4088 size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4089 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4092 if ((size == isize) &&
4093 !ext4_fs_is_busy(sbi) &&
4094 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4095 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4099 if (sbi->s_mb_group_prealloc <= 0) {
4100 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4104 /* don't use group allocation for large files */
4105 size = max(size, isize);
4106 if (size > sbi->s_mb_stream_request) {
4107 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4111 BUG_ON(ac->ac_lg != NULL);
4113 * locality group prealloc space are per cpu. The reason for having
4114 * per cpu locality group is to reduce the contention between block
4115 * request from multiple CPUs.
4117 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
4119 /* we're going to use group allocation */
4120 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4122 /* serialize all allocations in the group */
4123 mutex_lock(&ac->ac_lg->lg_mutex);
4126 static noinline_for_stack int
4127 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4128 struct ext4_allocation_request *ar)
4130 struct super_block *sb = ar->inode->i_sb;
4131 struct ext4_sb_info *sbi = EXT4_SB(sb);
4132 struct ext4_super_block *es = sbi->s_es;
4136 ext4_grpblk_t block;
4138 /* we can't allocate > group size */
4141 /* just a dirty hack to filter too big requests */
4142 if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4143 len = EXT4_CLUSTERS_PER_GROUP(sb);
4145 /* start searching from the goal */
4147 if (goal < le32_to_cpu(es->s_first_data_block) ||
4148 goal >= ext4_blocks_count(es))
4149 goal = le32_to_cpu(es->s_first_data_block);
4150 ext4_get_group_no_and_offset(sb, goal, &group, &block);
4152 /* set up allocation goals */
4153 ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4154 ac->ac_status = AC_STATUS_CONTINUE;
4156 ac->ac_inode = ar->inode;
4157 ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4158 ac->ac_o_ex.fe_group = group;
4159 ac->ac_o_ex.fe_start = block;
4160 ac->ac_o_ex.fe_len = len;
4161 ac->ac_g_ex = ac->ac_o_ex;
4162 ac->ac_flags = ar->flags;
4164 /* we have to define context: we'll we work with a file or
4165 * locality group. this is a policy, actually */
4166 ext4_mb_group_or_file(ac);
4168 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4169 "left: %u/%u, right %u/%u to %swritable\n",
4170 (unsigned) ar->len, (unsigned) ar->logical,
4171 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4172 (unsigned) ar->lleft, (unsigned) ar->pleft,
4173 (unsigned) ar->lright, (unsigned) ar->pright,
4174 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4179 static noinline_for_stack void
4180 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4181 struct ext4_locality_group *lg,
4182 int order, int total_entries)
4184 ext4_group_t group = 0;
4185 struct ext4_buddy e4b;
4186 struct list_head discard_list;
4187 struct ext4_prealloc_space *pa, *tmp;
4189 mb_debug(1, "discard locality group preallocation\n");
4191 INIT_LIST_HEAD(&discard_list);
4193 spin_lock(&lg->lg_prealloc_lock);
4194 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4196 spin_lock(&pa->pa_lock);
4197 if (atomic_read(&pa->pa_count)) {
4199 * This is the pa that we just used
4200 * for block allocation. So don't
4203 spin_unlock(&pa->pa_lock);
4206 if (pa->pa_deleted) {
4207 spin_unlock(&pa->pa_lock);
4210 /* only lg prealloc space */
4211 BUG_ON(pa->pa_type != MB_GROUP_PA);
4213 /* seems this one can be freed ... */
4215 spin_unlock(&pa->pa_lock);
4217 list_del_rcu(&pa->pa_inode_list);
4218 list_add(&pa->u.pa_tmp_list, &discard_list);
4221 if (total_entries <= 5) {
4223 * we want to keep only 5 entries
4224 * allowing it to grow to 8. This
4225 * mak sure we don't call discard
4226 * soon for this list.
4231 spin_unlock(&lg->lg_prealloc_lock);
4233 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4235 group = ext4_get_group_number(sb, pa->pa_pstart);
4236 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4237 ext4_error(sb, "Error loading buddy information for %u",
4241 ext4_lock_group(sb, group);
4242 list_del(&pa->pa_group_list);
4243 ext4_mb_release_group_pa(&e4b, pa);
4244 ext4_unlock_group(sb, group);
4246 ext4_mb_unload_buddy(&e4b);
4247 list_del(&pa->u.pa_tmp_list);
4248 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4253 * We have incremented pa_count. So it cannot be freed at this
4254 * point. Also we hold lg_mutex. So no parallel allocation is
4255 * possible from this lg. That means pa_free cannot be updated.
4257 * A parallel ext4_mb_discard_group_preallocations is possible.
4258 * which can cause the lg_prealloc_list to be updated.
4261 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4263 int order, added = 0, lg_prealloc_count = 1;
4264 struct super_block *sb = ac->ac_sb;
4265 struct ext4_locality_group *lg = ac->ac_lg;
4266 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4268 order = fls(pa->pa_free) - 1;
4269 if (order > PREALLOC_TB_SIZE - 1)
4270 /* The max size of hash table is PREALLOC_TB_SIZE */
4271 order = PREALLOC_TB_SIZE - 1;
4272 /* Add the prealloc space to lg */
4273 spin_lock(&lg->lg_prealloc_lock);
4274 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4276 spin_lock(&tmp_pa->pa_lock);
4277 if (tmp_pa->pa_deleted) {
4278 spin_unlock(&tmp_pa->pa_lock);
4281 if (!added && pa->pa_free < tmp_pa->pa_free) {
4282 /* Add to the tail of the previous entry */
4283 list_add_tail_rcu(&pa->pa_inode_list,
4284 &tmp_pa->pa_inode_list);
4287 * we want to count the total
4288 * number of entries in the list
4291 spin_unlock(&tmp_pa->pa_lock);
4292 lg_prealloc_count++;
4295 list_add_tail_rcu(&pa->pa_inode_list,
4296 &lg->lg_prealloc_list[order]);
4297 spin_unlock(&lg->lg_prealloc_lock);
4299 /* Now trim the list to be not more than 8 elements */
4300 if (lg_prealloc_count > 8) {
4301 ext4_mb_discard_lg_preallocations(sb, lg,
4302 order, lg_prealloc_count);
4309 * release all resource we used in allocation
4311 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4313 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4314 struct ext4_prealloc_space *pa = ac->ac_pa;
4316 if (pa->pa_type == MB_GROUP_PA) {
4317 /* see comment in ext4_mb_use_group_pa() */
4318 spin_lock(&pa->pa_lock);
4319 pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4320 pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4321 pa->pa_free -= ac->ac_b_ex.fe_len;
4322 pa->pa_len -= ac->ac_b_ex.fe_len;
4323 spin_unlock(&pa->pa_lock);
4328 * We want to add the pa to the right bucket.
4329 * Remove it from the list and while adding
4330 * make sure the list to which we are adding
4333 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4334 spin_lock(pa->pa_obj_lock);
4335 list_del_rcu(&pa->pa_inode_list);
4336 spin_unlock(pa->pa_obj_lock);
4337 ext4_mb_add_n_trim(ac);
4339 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4341 if (ac->ac_bitmap_page)
4342 page_cache_release(ac->ac_bitmap_page);
4343 if (ac->ac_buddy_page)
4344 page_cache_release(ac->ac_buddy_page);
4345 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4346 mutex_unlock(&ac->ac_lg->lg_mutex);
4347 ext4_mb_collect_stats(ac);
4351 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4353 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4357 trace_ext4_mb_discard_preallocations(sb, needed);
4358 for (i = 0; i < ngroups && needed > 0; i++) {
4359 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4368 * Main entry point into mballoc to allocate blocks
4369 * it tries to use preallocation first, then falls back
4370 * to usual allocation
4372 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4373 struct ext4_allocation_request *ar, int *errp)
4376 struct ext4_allocation_context *ac = NULL;
4377 struct ext4_sb_info *sbi;
4378 struct super_block *sb;
4379 ext4_fsblk_t block = 0;
4380 unsigned int inquota = 0;
4381 unsigned int reserv_clstrs = 0;
4384 sb = ar->inode->i_sb;
4387 trace_ext4_request_blocks(ar);
4389 /* Allow to use superuser reservation for quota file */
4390 if (IS_NOQUOTA(ar->inode))
4391 ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4394 * For delayed allocation, we could skip the ENOSPC and
4395 * EDQUOT check, as blocks and quotas have been already
4396 * reserved when data being copied into pagecache.
4398 if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4399 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4401 /* Without delayed allocation we need to verify
4402 * there is enough free blocks to do block allocation
4403 * and verify allocation doesn't exceed the quota limits.
4406 ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4408 /* let others to free the space */
4410 ar->len = ar->len >> 1;
4416 reserv_clstrs = ar->len;
4417 if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4418 dquot_alloc_block_nofail(ar->inode,
4419 EXT4_C2B(sbi, ar->len));
4422 dquot_alloc_block(ar->inode,
4423 EXT4_C2B(sbi, ar->len))) {
4425 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4436 ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4443 *errp = ext4_mb_initialize_context(ac, ar);
4449 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4450 if (!ext4_mb_use_preallocated(ac)) {
4451 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4452 ext4_mb_normalize_request(ac, ar);
4454 /* allocate space in core */
4455 *errp = ext4_mb_regular_allocator(ac);
4457 goto discard_and_exit;
4459 /* as we've just preallocated more space than
4460 * user requested originally, we store allocated
4461 * space in a special descriptor */
4462 if (ac->ac_status == AC_STATUS_FOUND &&
4463 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4464 *errp = ext4_mb_new_preallocation(ac);
4467 ext4_discard_allocated_blocks(ac);
4471 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4472 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4473 if (*errp == -EAGAIN) {
4475 * drop the reference that we took
4476 * in ext4_mb_use_best_found
4478 ext4_mb_release_context(ac);
4479 ac->ac_b_ex.fe_group = 0;
4480 ac->ac_b_ex.fe_start = 0;
4481 ac->ac_b_ex.fe_len = 0;
4482 ac->ac_status = AC_STATUS_CONTINUE;
4485 ext4_discard_allocated_blocks(ac);
4488 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4489 ar->len = ac->ac_b_ex.fe_len;
4492 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4500 ac->ac_b_ex.fe_len = 0;
4502 ext4_mb_show_ac(ac);
4504 ext4_mb_release_context(ac);
4507 kmem_cache_free(ext4_ac_cachep, ac);
4508 if (inquota && ar->len < inquota)
4509 dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4511 if (!ext4_test_inode_state(ar->inode,
4512 EXT4_STATE_DELALLOC_RESERVED))
4513 /* release all the reserved blocks if non delalloc */
4514 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4518 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4524 * We can merge two free data extents only if the physical blocks
4525 * are contiguous, AND the extents were freed by the same transaction,
4526 * AND the blocks are associated with the same group.
4528 static int can_merge(struct ext4_free_data *entry1,
4529 struct ext4_free_data *entry2)
4531 if ((entry1->efd_tid == entry2->efd_tid) &&
4532 (entry1->efd_group == entry2->efd_group) &&
4533 ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4538 static noinline_for_stack int
4539 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4540 struct ext4_free_data *new_entry)
4542 ext4_group_t group = e4b->bd_group;
4543 ext4_grpblk_t cluster;
4544 struct ext4_free_data *entry;
4545 struct ext4_group_info *db = e4b->bd_info;
4546 struct super_block *sb = e4b->bd_sb;
4547 struct ext4_sb_info *sbi = EXT4_SB(sb);
4548 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4549 struct rb_node *parent = NULL, *new_node;
4551 BUG_ON(!ext4_handle_valid(handle));
4552 BUG_ON(e4b->bd_bitmap_page == NULL);
4553 BUG_ON(e4b->bd_buddy_page == NULL);
4555 new_node = &new_entry->efd_node;
4556 cluster = new_entry->efd_start_cluster;
4559 /* first free block exent. We need to
4560 protect buddy cache from being freed,
4561 * otherwise we'll refresh it from
4562 * on-disk bitmap and lose not-yet-available
4564 page_cache_get(e4b->bd_buddy_page);
4565 page_cache_get(e4b->bd_bitmap_page);
4569 entry = rb_entry(parent, struct ext4_free_data, efd_node);
4570 if (cluster < entry->efd_start_cluster)
4572 else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4573 n = &(*n)->rb_right;
4575 ext4_grp_locked_error(sb, group, 0,
4576 ext4_group_first_block_no(sb, group) +
4577 EXT4_C2B(sbi, cluster),
4578 "Block already on to-be-freed list");
4583 rb_link_node(new_node, parent, n);
4584 rb_insert_color(new_node, &db->bb_free_root);
4586 /* Now try to see the extent can be merged to left and right */
4587 node = rb_prev(new_node);
4589 entry = rb_entry(node, struct ext4_free_data, efd_node);
4590 if (can_merge(entry, new_entry) &&
4591 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4592 new_entry->efd_start_cluster = entry->efd_start_cluster;
4593 new_entry->efd_count += entry->efd_count;
4594 rb_erase(node, &(db->bb_free_root));
4595 kmem_cache_free(ext4_free_data_cachep, entry);
4599 node = rb_next(new_node);
4601 entry = rb_entry(node, struct ext4_free_data, efd_node);
4602 if (can_merge(new_entry, entry) &&
4603 ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4604 new_entry->efd_count += entry->efd_count;
4605 rb_erase(node, &(db->bb_free_root));
4606 kmem_cache_free(ext4_free_data_cachep, entry);
4609 /* Add the extent to transaction's private list */
4610 ext4_journal_callback_add(handle, ext4_free_data_callback,
4611 &new_entry->efd_jce);
4616 * ext4_free_blocks() -- Free given blocks and update quota
4617 * @handle: handle for this transaction
4619 * @block: start physical block to free
4620 * @count: number of blocks to count
4621 * @flags: flags used by ext4_free_blocks
4623 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4624 struct buffer_head *bh, ext4_fsblk_t block,
4625 unsigned long count, int flags)
4627 struct buffer_head *bitmap_bh = NULL;
4628 struct super_block *sb = inode->i_sb;
4629 struct ext4_group_desc *gdp;
4630 unsigned int overflow;
4632 struct buffer_head *gd_bh;
4633 ext4_group_t block_group;
4634 struct ext4_sb_info *sbi;
4635 struct ext4_inode_info *ei = EXT4_I(inode);
4636 struct ext4_buddy e4b;
4637 unsigned int count_clusters;
4644 BUG_ON(block != bh->b_blocknr);
4646 block = bh->b_blocknr;
4650 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4651 !ext4_data_block_valid(sbi, block, count)) {
4652 ext4_error(sb, "Freeing blocks not in datazone - "
4653 "block = %llu, count = %lu", block, count);
4657 ext4_debug("freeing block %llu\n", block);
4658 trace_ext4_free_blocks(inode, block, count, flags);
4660 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4661 struct buffer_head *tbh = bh;
4664 BUG_ON(bh && (count > 1));
4666 for (i = 0; i < count; i++) {
4669 tbh = sb_find_get_block(inode->i_sb,
4673 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4674 inode, tbh, block + i);
4679 * We need to make sure we don't reuse the freed block until
4680 * after the transaction is committed, which we can do by
4681 * treating the block as metadata, below. We make an
4682 * exception if the inode is to be written in writeback mode
4683 * since writeback mode has weak data consistency guarantees.
4685 if (!ext4_should_writeback_data(inode))
4686 flags |= EXT4_FREE_BLOCKS_METADATA;
4689 * If the extent to be freed does not begin on a cluster
4690 * boundary, we need to deal with partial clusters at the
4691 * beginning and end of the extent. Normally we will free
4692 * blocks at the beginning or the end unless we are explicitly
4693 * requested to avoid doing so.
4695 overflow = EXT4_PBLK_COFF(sbi, block);
4697 if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4698 overflow = sbi->s_cluster_ratio - overflow;
4700 if (count > overflow)
4709 overflow = EXT4_LBLK_COFF(sbi, count);
4711 if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4712 if (count > overflow)
4717 count += sbi->s_cluster_ratio - overflow;
4722 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4724 if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4725 ext4_get_group_info(sb, block_group))))
4729 * Check to see if we are freeing blocks across a group
4732 if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4733 overflow = EXT4_C2B(sbi, bit) + count -
4734 EXT4_BLOCKS_PER_GROUP(sb);
4737 count_clusters = EXT4_NUM_B2C(sbi, count);
4738 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4743 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4749 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4750 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4751 in_range(block, ext4_inode_table(sb, gdp),
4752 EXT4_SB(sb)->s_itb_per_group) ||
4753 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4754 EXT4_SB(sb)->s_itb_per_group)) {
4756 ext4_error(sb, "Freeing blocks in system zone - "
4757 "Block = %llu, count = %lu", block, count);
4758 /* err = 0. ext4_std_error should be a no op */
4762 BUFFER_TRACE(bitmap_bh, "getting write access");
4763 err = ext4_journal_get_write_access(handle, bitmap_bh);
4768 * We are about to modify some metadata. Call the journal APIs
4769 * to unshare ->b_data if a currently-committing transaction is
4772 BUFFER_TRACE(gd_bh, "get_write_access");
4773 err = ext4_journal_get_write_access(handle, gd_bh);
4776 #ifdef AGGRESSIVE_CHECK
4779 for (i = 0; i < count_clusters; i++)
4780 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4783 trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4785 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4789 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4790 struct ext4_free_data *new_entry;
4792 * blocks being freed are metadata. these blocks shouldn't
4793 * be used until this transaction is committed
4796 new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4799 * We use a retry loop because
4800 * ext4_free_blocks() is not allowed to fail.
4803 congestion_wait(BLK_RW_ASYNC, HZ/50);
4806 new_entry->efd_start_cluster = bit;
4807 new_entry->efd_group = block_group;
4808 new_entry->efd_count = count_clusters;
4809 new_entry->efd_tid = handle->h_transaction->t_tid;
4811 ext4_lock_group(sb, block_group);
4812 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4813 ext4_mb_free_metadata(handle, &e4b, new_entry);
4815 /* need to update group_info->bb_free and bitmap
4816 * with group lock held. generate_buddy look at
4817 * them with group lock_held
4819 if (test_opt(sb, DISCARD)) {
4820 err = ext4_issue_discard(sb, block_group, bit, count);
4821 if (err && err != -EOPNOTSUPP)
4822 ext4_msg(sb, KERN_WARNING, "discard request in"
4823 " group:%d block:%d count:%lu failed"
4824 " with %d", block_group, bit, count,
4827 EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4829 ext4_lock_group(sb, block_group);
4830 mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4831 mb_free_blocks(inode, &e4b, bit, count_clusters);
4834 ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4835 ext4_free_group_clusters_set(sb, gdp, ret);
4836 ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4837 ext4_group_desc_csum_set(sb, block_group, gdp);
4838 ext4_unlock_group(sb, block_group);
4840 if (sbi->s_log_groups_per_flex) {
4841 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4842 atomic64_add(count_clusters,
4843 &sbi->s_flex_groups[flex_group].free_clusters);
4846 if (flags & EXT4_FREE_BLOCKS_RESERVE && ei->i_reserved_data_blocks) {
4847 percpu_counter_add(&sbi->s_dirtyclusters_counter,
4849 spin_lock(&ei->i_block_reservation_lock);
4850 if (flags & EXT4_FREE_BLOCKS_METADATA)
4851 ei->i_reserved_meta_blocks += count_clusters;
4853 ei->i_reserved_data_blocks += count_clusters;
4854 spin_unlock(&ei->i_block_reservation_lock);
4855 if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4856 dquot_reclaim_block(inode,
4857 EXT4_C2B(sbi, count_clusters));
4858 } else if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4859 dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4860 percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4862 ext4_mb_unload_buddy(&e4b);
4864 /* We dirtied the bitmap block */
4865 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4866 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4868 /* And the group descriptor block */
4869 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4870 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4874 if (overflow && !err) {
4882 ext4_std_error(sb, err);
4887 * ext4_group_add_blocks() -- Add given blocks to an existing group
4888 * @handle: handle to this transaction
4890 * @block: start physical block to add to the block group
4891 * @count: number of blocks to free
4893 * This marks the blocks as free in the bitmap and buddy.
4895 int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4896 ext4_fsblk_t block, unsigned long count)
4898 struct buffer_head *bitmap_bh = NULL;
4899 struct buffer_head *gd_bh;
4900 ext4_group_t block_group;
4903 struct ext4_group_desc *desc;
4904 struct ext4_sb_info *sbi = EXT4_SB(sb);
4905 struct ext4_buddy e4b;
4906 int err = 0, ret, blk_free_count;
4907 ext4_grpblk_t blocks_freed;
4909 ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4914 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4916 * Check to see if we are freeing blocks across a group
4919 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4920 ext4_warning(sb, "too much blocks added to group %u\n",
4926 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4932 desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4938 if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4939 in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4940 in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4941 in_range(block + count - 1, ext4_inode_table(sb, desc),
4942 sbi->s_itb_per_group)) {
4943 ext4_error(sb, "Adding blocks in system zones - "
4944 "Block = %llu, count = %lu",
4950 BUFFER_TRACE(bitmap_bh, "getting write access");
4951 err = ext4_journal_get_write_access(handle, bitmap_bh);
4956 * We are about to modify some metadata. Call the journal APIs
4957 * to unshare ->b_data if a currently-committing transaction is
4960 BUFFER_TRACE(gd_bh, "get_write_access");
4961 err = ext4_journal_get_write_access(handle, gd_bh);
4965 for (i = 0, blocks_freed = 0; i < count; i++) {
4966 BUFFER_TRACE(bitmap_bh, "clear bit");
4967 if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4968 ext4_error(sb, "bit already cleared for block %llu",
4969 (ext4_fsblk_t)(block + i));
4970 BUFFER_TRACE(bitmap_bh, "bit already cleared");
4976 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4981 * need to update group_info->bb_free and bitmap
4982 * with group lock held. generate_buddy look at
4983 * them with group lock_held
4985 ext4_lock_group(sb, block_group);
4986 mb_clear_bits(bitmap_bh->b_data, bit, count);
4987 mb_free_blocks(NULL, &e4b, bit, count);
4988 blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4989 ext4_free_group_clusters_set(sb, desc, blk_free_count);
4990 ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
4991 ext4_group_desc_csum_set(sb, block_group, desc);
4992 ext4_unlock_group(sb, block_group);
4993 percpu_counter_add(&sbi->s_freeclusters_counter,
4994 EXT4_NUM_B2C(sbi, blocks_freed));
4996 if (sbi->s_log_groups_per_flex) {
4997 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4998 atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
4999 &sbi->s_flex_groups[flex_group].free_clusters);
5002 ext4_mb_unload_buddy(&e4b);
5004 /* We dirtied the bitmap block */
5005 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5006 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5008 /* And the group descriptor block */
5009 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5010 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5016 ext4_std_error(sb, err);
5021 * ext4_trim_extent -- function to TRIM one single free extent in the group
5022 * @sb: super block for the file system
5023 * @start: starting block of the free extent in the alloc. group
5024 * @count: number of blocks to TRIM
5025 * @group: alloc. group we are working with
5026 * @e4b: ext4 buddy for the group
5028 * Trim "count" blocks starting at "start" in the "group". To assure that no
5029 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5030 * be called with under the group lock.
5032 static int ext4_trim_extent(struct super_block *sb, int start, int count,
5033 ext4_group_t group, struct ext4_buddy *e4b)
5035 struct ext4_free_extent ex;
5038 trace_ext4_trim_extent(sb, group, start, count);
5040 assert_spin_locked(ext4_group_lock_ptr(sb, group));
5042 ex.fe_start = start;
5043 ex.fe_group = group;
5047 * Mark blocks used, so no one can reuse them while
5050 mb_mark_used(e4b, &ex);
5051 ext4_unlock_group(sb, group);
5052 ret = ext4_issue_discard(sb, group, start, count);
5053 ext4_lock_group(sb, group);
5054 mb_free_blocks(NULL, e4b, start, ex.fe_len);
5059 * ext4_trim_all_free -- function to trim all free space in alloc. group
5060 * @sb: super block for file system
5061 * @group: group to be trimmed
5062 * @start: first group block to examine
5063 * @max: last group block to examine
5064 * @minblocks: minimum extent block count
5066 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5067 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5071 * ext4_trim_all_free walks through group's block bitmap searching for free
5072 * extents. When the free extent is found, mark it as used in group buddy
5073 * bitmap. Then issue a TRIM command on this extent and free the extent in
5074 * the group buddy bitmap. This is done until whole group is scanned.
5076 static ext4_grpblk_t
5077 ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5078 ext4_grpblk_t start, ext4_grpblk_t max,
5079 ext4_grpblk_t minblocks)
5082 ext4_grpblk_t next, count = 0, free_count = 0;
5083 struct ext4_buddy e4b;
5086 trace_ext4_trim_all_free(sb, group, start, max);
5088 ret = ext4_mb_load_buddy(sb, group, &e4b);
5090 ext4_error(sb, "Error in loading buddy "
5091 "information for %u", group);
5094 bitmap = e4b.bd_bitmap;
5096 ext4_lock_group(sb, group);
5097 if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5098 minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5101 start = (e4b.bd_info->bb_first_free > start) ?
5102 e4b.bd_info->bb_first_free : start;
5104 while (start <= max) {
5105 start = mb_find_next_zero_bit(bitmap, max + 1, start);
5108 next = mb_find_next_bit(bitmap, max + 1, start);
5110 if ((next - start) >= minblocks) {
5111 ret = ext4_trim_extent(sb, start,
5112 next - start, group, &e4b);
5113 if (ret && ret != -EOPNOTSUPP)
5116 count += next - start;
5118 free_count += next - start;
5121 if (fatal_signal_pending(current)) {
5122 count = -ERESTARTSYS;
5126 if (need_resched()) {
5127 ext4_unlock_group(sb, group);
5129 ext4_lock_group(sb, group);
5132 if ((e4b.bd_info->bb_free - free_count) < minblocks)
5138 EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5141 ext4_unlock_group(sb, group);
5142 ext4_mb_unload_buddy(&e4b);
5144 ext4_debug("trimmed %d blocks in the group %d\n",
5151 * ext4_trim_fs() -- trim ioctl handle function
5152 * @sb: superblock for filesystem
5153 * @range: fstrim_range structure
5155 * start: First Byte to trim
5156 * len: number of Bytes to trim from start
5157 * minlen: minimum extent length in Bytes
5158 * ext4_trim_fs goes through all allocation groups containing Bytes from
5159 * start to start+len. For each such a group ext4_trim_all_free function
5160 * is invoked to trim all free space.
5162 int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5164 struct ext4_group_info *grp;
5165 ext4_group_t group, first_group, last_group;
5166 ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5167 uint64_t start, end, minlen, trimmed = 0;
5168 ext4_fsblk_t first_data_blk =
5169 le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5170 ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5173 start = range->start >> sb->s_blocksize_bits;
5174 end = start + (range->len >> sb->s_blocksize_bits) - 1;
5175 minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5176 range->minlen >> sb->s_blocksize_bits);
5178 if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5179 start >= max_blks ||
5180 range->len < sb->s_blocksize)
5182 if (end >= max_blks)
5184 if (end <= first_data_blk)
5186 if (start < first_data_blk)
5187 start = first_data_blk;
5189 /* Determine first and last group to examine based on start and end */
5190 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5191 &first_group, &first_cluster);
5192 ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5193 &last_group, &last_cluster);
5195 /* end now represents the last cluster to discard in this group */
5196 end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5198 for (group = first_group; group <= last_group; group++) {
5199 grp = ext4_get_group_info(sb, group);
5200 /* We only do this if the grp has never been initialized */
5201 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5202 ret = ext4_mb_init_group(sb, group);
5208 * For all the groups except the last one, last cluster will
5209 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5210 * change it for the last group, note that last_cluster is
5211 * already computed earlier by ext4_get_group_no_and_offset()
5213 if (group == last_group)
5216 if (grp->bb_free >= minlen) {
5217 cnt = ext4_trim_all_free(sb, group, first_cluster,
5227 * For every group except the first one, we are sure
5228 * that the first cluster to discard will be cluster #0.
5234 atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5237 range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;